HALCON 26.11 expands its AI object detection with oriented bounding boxes, enabling more precise localization of rotated, elongated, or closely spaced objects. By representing an object’s orientation in addition to its position and size, oriented bounding boxes reduce irrelevant background and enable orientation-aware evaluation.
Per-class confidence thresholds provide further control by allowing users to adjust the balance between detection rates and false positives individually for each object class.
For edge deployment, the new ONNX Runtime AI² interface enables more efficient deep learning inference on Linux-based edge devices using the 64-bit Arm architecture.
HALCON 26.11 extends automatic contour optimization to slightly deformable objects. This makes it easier to create robust Shape Matching models for real-world parts that slightly differ in every image. For example, metal or plastic components may vary slightly because of production tolerances, perspective effects, or minor deformations.
Previously, such variations could limit automatic contour optimization and require users to disable it, manually edit model contours, or carefully select training images with minimal geometric differences. HALCON 26.11 can now account for expected local shape variations while identifying and removing genuinely unstable contours. Using multiple representative training images, users can create compact and robust models with less manual contour editing and apply automatic contour optimization to a broader range of real-world matching tasks.
HALCON 26.11 simplifies common 3D vision tasks and delivers more reliable results from real-world data. When fitting geometric shapes, users can now incorporate known object properties such as size or orientation. This helps reduce ambiguities, for example when objects are only partially visible or measurement data contains noise.
Additional improvements make everyday 3D processing workflows easier to handle. A new operator simplifies the calculation of intersections between lines and planes. HALCON 26.11 also preserves neighborhood information when sampling XYZ-mapped 3D data, provides more consistent handling of XYZ mappings when object models are merged, and gives users greater control over surface-normal calculation.
Cybersecurity and regulatory compliance are increasingly important throughout the lifecycle of connected products. With HALCON 26.11, MVTec strengthens the processes supporting vulnerability handling and security updates. Together with the Software Bills of Materials already provided for HALCON, these measures give customers transparent information about included software components and a clearer basis for addressing security issues.
This helps manufacturers assess dependencies, maintain their products, and prepare relevant documentation with less effort. HALCON thus supports manufacturers in addressing EU Cyber Resilience Act requirements, while responsibility for assessing the compliance of the complete final product remains with its manufacturer.
With HALCON 26.05, automatic contour optimization for shape-based matching (SBM) enables users to automatically remove unstable or misleading contours from SBM models. Reflections, shadows, or random texture often introduce unreliable contours that reduce matching robustness and require labor-intensive manual cleanup. With the new feature, users simply provide sample images of real object instances. Based on these samples, the system analyzes which contours appear consistently across variations and retains only the stable contours while removing unreliable ones.
By focusing the model on robust contours only, matching becomes faster, more stable, and more accurate. The feature replaces time-consuming manual contour editing with a data-driven optimization step during model training. It is particularly valuable for reflective or textured objects, such as mechanical or electronic components, and for automation scenarios like feeder-based pick-and-place systems where stable matching is critical. The automated optimization also reduces the effort for retraining models when new objects are introduced, making inline retooling more practical in production environments.
HALCON 26.05 expands its code reader with Data Matrix rectification, enabling reliable reading of Data Matrix codes even when they appear on curved or deformed surfaces. In many industrial applications, codes are printed on non-flat materials, which can distort the symbol geometry and reduce reading reliability. With the new rectification capability, HALCON compensates for these distortions before decoding, significantly improving robustness in such scenarios.
The rectification step can be enabled optionally within the code reader and integrates seamlessly into existing workflows. Although processing time is slightly higher than with standard Data Matrix reading, the improved decoding reliability enables stable operation in demanding environments. Typical applications include codes printed on cylindrical components, curved packaging, or flexible materials used in manufacturing and logistics.
HALCON 26.05 introduces enhanced data augmentation for deep-learning workflows. The new approach replaces the previous procedure-based augmentation and preprocessing steps with configurable operators that integrate directly into HALCON deep-learning pipelines.
Users can define augmentation pipelines programmatically, apply transformation techniques such as geometric transforms, color variations, and blurring, and preview the resulting image variations. This lets developers test and refine augmentation strategies quickly within their existing workflow. Training samples generated this way help model robustness and generalization, which can reduce the reliance on large training datasets.
This feature is initially available for object detection and instance segmentation, supporting more reliable training and improving model generalization under challenging conditions such as varying illumination, perspective changes, occlusions, or noisy image data.
A new generation of deep-learning-based object detection is available with HALCON 26.05, delivering up to 5x faster inference while maintaining high detection accuracy. The new architecture enables efficient detection of objects and is optimized for demanding machine vision scenarios where both speed and precision are critical.
Users can train and run detection models directly within HALCON and start from MVTec-provided pretrained models that can be adapted to specific applications. The anchor-free detection approach improves bounding-box localization and performs reliably even for small objects and varying object sizes. Integrated data augmentation techniques further increase robustness against changes in illumination, rotation, distortion, and partial occlusion. The feature integrates directly into HALCON workflows and supports inspection, localization, and sorting tasks across industrial automation applications.
HALCON 25.11 introduces Continual Learning – Classification, a new technology that makes training and maintaining classification models faster and more flexible. Users can create models with only few images per class and adapt them at any time – for example, to refine existing classes or add new ones.
Unlike conventional deep learning, this approach prevents catastrophic forgetting and keeps maintenance effort low. Based on MVTec’s pretrained models optimized for industrial scenarios, applications can be updated quickly without full retraining. Because the method requires minimal computing power, updates can even be performed directly on edge devices, eliminating the need for external training hardware while ensuring efficient, long-term operation.
The result is a flexible solution that evolves with changing production conditions and remains suitable for embedded and edge environments such as smart cameras, sensors, and inspection modules.
With Score Visualization for Shape Matching in HALCON 25.11, users gain increased transparency when setting up shape matching applications. Instead of only returning an overall score, the feature provides a breakdown of how different model parts contribute to the final result. By configuring color-coded bins, users can immediately see which areas match well and which perform poorly, for example due to shadows or unwanted textures. This visual feedback makes it much easier to refine models, remove problematic parts, and optimize applications – a major usability advantage especially for non-expert users.
The feature can also support advanced scenarios in robotics, helping determine which object in a stack is least covered and should be picked first.
With new Deep OCR recognition models in HALCON 25.11, text reading becomes faster and more resource-efficient without compromising accuracy. The models deliver up to 50× faster inference on embedded devices. All models are pretrained by MVTec on industrial image data, and include the proven alignment preprocessing, which improves recognition when text varies in position or orientation. Thanks to their optimized architecture, they enable real-time OCR applications on low-power devices while maintaining high accuracy. This makes the models ideal for demanding inline applications such as serial number inspection, label verification, or lot tracking OCR tasks, across industries from logistics and packaging to pharmaceuticals, consumer goods, and medical technology.
With HALCON 25.11, MVTec adds support for the MobileNetV4 series, an efficient new generation of deep learning models optimized for resource-constrained systems and edge devices. These models support both classification and object detection tasks and deliver high accuracy while maintaining low computational requirements. Users benefit from fast inference times, lower system costs, and straightforward integration into existing HALCON projects. All models are pretrained by MVTec, ensuring strong performance for various downstream tasks such as quality inspection, product classification, presence detection, and surface defect analysis. Typical industries include automation, electronics, packaging, food, and medical technology.
With HALCON 25.11, code reading and print quality inspection (PQI) become even more robust and versatile.
QR code detection has been improved for challenging cases such as curved or deformed surfaces. A more powerful candidate search significantly raises the detection rate, while runtime has been reduced for standard scenarios – enabling reliable reading in industries like logistics, packaging, food production, and bottle labeling.
The bar code reader has also been enhanced for Code 128 and GS1-128, making it more tolerant to irregular bar widths caused by printing variations or local distortions. This increases decoding reliability across diverse industrial applications.
In addition, HALCON now supports the latest print quality inspection standards ISO/IEC 15415:2024 and ISO/IEC 29158:2025. This ensures code quality can be verified according to the most up-to-date requirements in sectors such as pharmaceuticals, food, and logistics.
Together, these enhancements provide compliance, long-term process stability, and higher robustness across a wide range of industrial code reading applications.
With HALCON 25.11, MVTec provides Software Bills of Materials (SBOMs), giving users transparent insight into the software components included in the product. SBOMs are becoming a key requirement under new regulations such as the EU Cyber Resilience Act and are increasingly demanded in process- and safety-critical industries.
By providing SBOMs directly with HALCON, MVTec simplifies compliance and reduces workload for customers. Delivered as machine-readable SPDX JSON files, SBOMs make it easier to perform vulnerability and license analyses, fulfill regulatory obligations, and react quickly to newly discovered risks. The result is less integration effort, lower long-term costs, and greater confidence in meeting both regulatory and customer requirements.
For the release notes of HDevelopEVO and HALCON Script Engine (HScriptEngine), see this separate page .
All HALCON 26.05 Progress-Steady licenses or licenses of earlier versions must be replaced or upgraded. Please contact your local sales partner. HALCON 26.11.1.0 Steady licenses will be downwards compatible to HALCON 26.05 Progress-Steady.
Compared to HALCON 26.05 Progress, many extensions have been introduced. Thus, the HALCON 26.11.1.0 Steady libraries are not binary compatible with HALCON 26.05 Progress or earlier versions. However, HALCON 26.11.1.0 Steady is mostly source-code compatible to HALCON 26.05 Progress except for the changes listed below:
In previous versions, tuple_rand(0) threw an error. Starting with HALCON 26.11, it returns an empty tuple instead.
More information.union_object_model_3d for Method set to 'points_surface' does not return mappings anymore. If mappings are wanted for further processing, please set Method to 'points_surface_single_viewpoint'.
More information.sample_object_model_3d with Method set to 'xyz_mapping' or 'xyz_mapping_compute_normals' now stores a zoomed, dense output XYZ mapping and does not copy sparse values from the input XYZ mapping anymore.
More information.Due to a code refactoring, extension packages using the functions HRLClip2, HRLRelClip, or HRLClip are no longer source-code compatible and must be adapted.
More information.The Postscript window type has been removed.
The fallback mode that allowed HALCON to open windows with reduced functionality if the hcanvas library could not be loaded has been removed.
More information.The default memory allocator on Windows systems is no longer mimalloc, but the system allocator. This is the same behavior as on Linux systems. You can use set_system('memory_allocator', 'mimalloc'), to re-enable the mimalloc allocator.
More information.Some OBJ files previously written by HALCON were not valid OBJ files. The meshes' vertex normals are lost when reading OBJ files written by previous versions of HALCON.
More information.In earlier versions of HALCON, the GenParamName and GenParamValue parameters of enqueue_message were simply ignored. Now, you must pass an empty tuple for GenParamName and GenParamValue if you do not want to pass any valid generic parameters to enqueue_message.
More information.As the behavior of linking edge points to contours changed in edges_sub_pix and edges_color_sub_pix, different results are to be expected. This can also result in weak edges being omitted from the returned contour, which were present before.
More information.Internal optimizations in the ONNX reader may change the ordering of output dimensions for models loaded from ONNX files compared to previous versions of HALCON. Customers who use ONNX models (including YOLO models) in their programs may need to update their integration accordingly.
More information.HALCON now uses OpenJPEG to read JPEG 2000 images. While this should be faster and more robust than the previously used libjasper, there will be differences in compression between images written by both libraries.
More information.Applications setting the parameter 'contrast_min' for the (Micro) QR Code reader must be adapted, as this parameter has been removed and set_data_code_2d_param and get_data_code_2d_param now return an error with this parameter for these readers. Applications that explicitly trained 'contrast' with the (Micro) QR Code reader must be adapted as well.
More information.For Deep 3D Matching, the parameter 'sort_single_views_last' is by default set to 'true' when reading models created with previous HALCON versions. This may lead to a different ordering of the results when setting 'min_num_views' to 1 and having more than one camera.
More information.Legacy handle mode has been removed.
HALCON now implements the new version of the standard ISO/IEC 15415, namely ISO/IEC15415:2024, such that the output of the operator get_data_code_2d_results with the parameters 'quality_isoiec15415' resp. 'quality_isoiec15415_*' is not compatible with previous HALCON versions:
For the PDF417 reader, the parameter 'quality_isoiec15415_float_grades' of the operator get_data_code_2d_results is no longer available, as the grades are always continuous in the new HALCON version.
More information.The update of the NVIDIA CUDA libraries significantly reduces the memory footprint in most cases. However, on some cards there may be performance regressions. The new set_system parameter 'cudnn_tuning' can be used to partially mitigate these issues, for example, by setting it to 'fastest'.
The TensorRT AI² interface now requires compute capability 7.5. As a consequence, older NVIDIA architectures up to the Pascal architecture (this includes the GeForce GTX 10xx series) are not supported with the TensorRT AI² Interface anymore. See https://docs.nvidia.com/deeplearning/tensorrt/10.12.0/getting-started/support-matrix.html.
More information.For detector models retrieved from a Deep 3D Matching model, normalization and augmentation are now carried out within train_dl_model_batch. All training datasets previously preprocessed have to be preprocessed again in preprocess_dl_dataset to be compatible.
Note that these changes do not affect standard detector models, which are not used in Deep 3D Matching.
The loss values and gradients of the CTC loss layer of the DL Framework are now normalized by the respective target sequence lengths. To approximate the behavior of previous versions, it may be necessary to multiply the learning rate by the mean target length.
More information.HALCON now implements the new version of the standard ISO/IEC 29158, namely ISO/IEC 29158:2025. Consequently, the output of get_data_code_2d_results with the parameters 'quality_isoiec29158' resp. 'quality_isoiec29158_*' is no longer compatible with previous HALCON versions:
The following parameters have been entirely removed from 'get_data_code_2d_results' and are no longer available:
uncalibrated_photometric_stereo now raises an error if ResultType is not set to 'all'.
More information.In very rare cases, a bug fix in an internal function can change the behavior of multiple operators, including but not limited to gen_region_polygon_filled and gray_inside.
Please re-compile all C, C++, or .NET programs developed with HALCON 26.05 Progress. The incompatibility with HALCON 26.05 Progress or earlier versions mainly concerns the binaries, with only few changes in the language interfaces. If you encounter problems during recompiling your programs, please check the detailed description of changes below.
In general, HALCON 26.11.1.0 Steady and HALCON 26.05 Progress image acquisition interfaces are library compatible.
HALCON 26.11.1.0 Steady includes only a subset of available image acquisition interfaces. You can download additional interfaces from our web server.
In general, HALCON 26.11.1.0 Steady and HALCON 26.05 Progress digital I/O interfaces are library compatible.
HALCON 26.11.1.0 Steady includes only a subset of available digital I/O interfaces. You can download additional interfaces from our web server.
Please re-generate your own extension packages developed with HALCON 26.05 Progress.
On January 19 2038 at 03:14:07 UTC, the 32-bit time_t values used by the armv7a-linux version of HALCON will overflow, causing HALCON (and any other software using 32-bit time_t values) to stop working correctly (see https://en.wikipedia.org/wiki/Year_2038_problem for more details). To mitigate this issue, MVTec can provide an armv7a-linux-glibc2_34 variant of HALCON that uses 64-bit time_t values, which will not overflow for billions of years to come. However, this variant requires kernel version 5.6 or later and glibc version 2.34 or later. Apart from 64-bit time_t support and the modified system requirements, the armv7a-linux-glibc2_34 variant of HALCON is functionally identical to the armv7a-linux variant.
The following functionality may be discontinued in a future major release:
See the reference manual entries of legacy operators for details on how to replace them.
The following legacy operators have been removed from HALCON (Legacy/Matching, Component-Based):
clear_all_component_models, clear_all_training_components, clear_component_model, clear_training_components, cluster_model_components, create_component_model, create_trained_component_model, deserialize_component_model, deserialize_training_components, find_component_model, gen_initial_components, get_component_model_params, get_component_model_tree, get_component_relations, get_found_component_model, get_training_components, inspect_clustered_components, modify_component_relations, read_component_model, read_training_components, serialize_component_model, serialize_training_components, train_model_components, write_component_model, write_training_components
The following classes have been removed from HALCON:
HComponentModel, HComponentTraining
Official support for Visual Studio 2013 has been discontinued. Please use Visual Studio 2017 or later for compiling HALCON applications.
You will still need Visual Studio 2022 or later to use the HALCON Variable Inspect Visual Studio Extension as announced separately.
The support for the 32‑bit HALCON Variable Inspect Visual Studio Extension has been removed. Please switch to the 64‑bit versions of Microsoft Visual Studio 2022 or 2026.
The PostScript window type has been removed.
The fallback mode that allowed HALCON to open windows with reduced functionality if the hcanvas library could not be loaded has been removed.
The following deprecated operators have been removed from HALCON:
disp_lut, get_fixed_lut, get_lut_style, set_fixed_lut, set_lut_style, write_lut, open_textwindow, clear_rectangle, move_rectangle, slide_image, disp_distribution, get_tshape, query_tshape, set_tshape, get_comprise, set_comprise, get_fix, set_fix, set_line_approx, get_line_approx, get_pixel, set_pixel
The legacy functionality Sample Based Identification has been removed.
Support for the 32‑bit HALCON Variable Inspect Visual Studio Extension has been moved to legacy status. To ensure continued support and compatibility, please switch to the 64‑bit versions of Microsoft Visual Studio 2022 or 2026.
The Hailo AI² plugin has been transitioned to legacy status and is no longer actively developed or maintained.
The legacy operators and procedures listed below, which used the legacy dl_classifier handle, are not supported any longer.
Operators:
Procedures:
The legacy handle mode has been announced legacy since HALCON 24.11. The mode shifts the ownership of a handle to the user via the parameter 'legacy_handle_mode' of the set_system operator, and has been removed from HALCON.
The following deprecated operators have been removed from HALCON:
clear_all_bar_code_models, clear_all_barriers, clear_all_calib_data, clear_all_camera_setup_models, clear_all_class_gmm, clear_all_class_knn, clear_all_class_lut, clear_all_class_mlp, clear_all_class_svm, clear_all_class_train_data, clear_all_color_trans_luts, clear_all_component_models, clear_all_conditions, clear_all_data_code_2d_models, clear_all_deformable_models, clear_all_descriptor_models, clear_all_events, clear_all_lexica, clear_all_matrices, clear_all_metrology_models, clear_all_mutexes, clear_all_ncc_models, clear_all_object_model_3d, clear_all_ocr_class_knn, clear_all_ocr_class_mlp, clear_all_ocr_class_svm, clear_all_sample_identifiers, clear_all_scattered_data_interpolators, clear_all_serialized_items, clear_all_shape_model_3d, clear_all_shape_models, clear_all_sheet_of_light_models, clear_all_stereo_models, clear_all_surface_matching_results, clear_all_surface_models, clear_all_templates, clear_all_text_models, clear_all_text_results, clear_all_training_components, clear_all_variation_models,
close_all_bg_esti, close_all_class_box, close_all_files, close_all_framegrabbers, close_all_measures, close_all_ocrs, close_all_ocvs, close_all_serials, close_all_sockets
The following legacy operators have been removed from HALCON:
Legacy/Classification
Legacy/Segmentation
Legacy/Tools
The following classes have been removed from HALCON:
The following legacy operators have been removed from HALCON:
The following classes have been removed from HALCON:
The legacy operators set_insert, get_insert, and query_insert have been removed.
As announced in previous releases, MVTec has discontinued the support of ARM 32-bit systems with the release of HALCON 25.11 Progress. We recommend switching to ARM 64-bit platforms for future applications.
HALCON 26.11.1.0 Steady has been compiled for the x64-win64 platform version for Windows 10 (x64 editions), 11, Windows Server 2016, 2019, 2022, 2025 on Intel 64 or AMD 64 with SSE2 (AVX2 dispatch) processors.
HALCON 26.11.1.0 Steady has been compiled for the following Linux platform versions:
Please refer to the Installation Guide for detailed system requirements corresponding to the different Application Binary Interfaces.
The following issues are relevant in the context of cybersecurity:
Internal security checks identified the security vulnerabilities CVE-2026-27489, CVE-2026-34447, CVE-2026-34446. These vulnerabilities concern path traversal vulnerabilities in ONNX. These vulnerabilities have been addressed by updating the ONNX third-party library to version 1.22.0, which is no longer affected by them.
Internal security checks identified the security vulnerability CVE-2026-34445. The vulnerability concerns the ONNX Python interface, which is not being used. While this CVE is not applicable to our product and has never affected HALCON, it has also been addressed by updating the ONNX third-party library to version 1.22.0, which is no longer affected by it.
Internal security checks identified the security vulnerability CVE-2021-4048. The vulnerability concerns the CLARRV, DLARRV, SLARRV, and ZLARRV functions in lapack. HALCON does neither ship nor use these functions. Exploitation in the context of HALCON is therefore not possible and the CVE is not applicable for our product.
Internal security checks identified the security vulnerability CVE-2026-44512. This vulnerability could cause an unrecoverable denial of service when processing an untrusted model. While this CVE is not applicable to our product and has never affected HALCON, it has also been addressed by updating the ONNX third-party library to version 1.22.0, which is no longer affected by it.
The third-party library GenApi has been updated to version 3.5.1. This fixes various security vulnerabilities by updating the third-party library Expat to version 2.8.1.
Internal security checks identified the security vulnerability CVE-2026-63632. While this CVE is not applicable to our product and has never affected HALCON, it has also been addressed by updating the ONNX third-party library to version 1.22.0, which is no longer affected by it.
The third-party library libxml2 has been updated to version 2.15.4. This fixes the security vulnerabilities CVE-2026-86137, CVE-2026-86138, CVE-2026-86139, CVE-2026-86140, CVE-2026-86141, CVE-2026-86142, CVE-2026-86143, and CVE-2026-86144.
Internal security checks identified the security vulnerabilities CVE-2026-89157, CVE-2026-89161, CVE-2026-89158, CVE-2026-89160, and CVE-2026-89156. The vulnerabilities concern specific functions and settings of the pcre2 library. These functions and settings are not used and the CVEs are not applicable to our product and do not affect HALCON.
Internal security checks identified the security vulnerability CVE-2026-85091 (third-party library zlib). This vulnerability does not apply to HALCON, as the vulnerable APIs are not used in HALCON.
The third-party library OpenSSL has been updated to version 3.5.9. This fixes numerous security vulnerabilities.
The third-party library libpng has been updated to 1.6.59. This fixes the security vulnerability CVE-2026-46675.
Internal security checks identified the security vulnerabilities CVE-2024-0102, CVE-2024-0072, CVE-2023-25523, CVE-2024-0076 (third-party CUDA SDK, cublas). The vulnerabilities concern the tools nvdisasm and cuobjdump, which are shipped alongside the CUDA SDK. HALCON neither ships nor uses these binaries. Exploitation in the context of HALCON is therefore not possible, and the CVEs are not applicable for our product.
The third-party library Apache Thrift has been updated to version 0.24.0 to provide a fix for CVE-2026-55971, CVE-2026-58662, CVE-2026-48145, CVE-2026-48586, CVE-2026-55969, and CVE-2026-55970. This library is used for the communication between HDevelop and HDevEngine during remote debugging. HALCON applications that do not use HDevEngine are unaffected.
HDevEngine production applications that run without enabling the debug server are also unaffected. If the debug server was enabled, exploiting potential vulnerabilities required malicious third-party software to be able to open a socket connection to the host/port combination configured for remote debugging.
Another path to exploiting potential vulnerabilities was to trick an HDevelop user into initiating a remote debugging connection to a malicious third-party server.
HALCON now uses version 8.41a of the CodeMeter Runtime. This update fixes security vulnerabilities in CodeMeter Runtime.
The third-party library OpenSSL has been updated to version 3.5.8. This fixes the security vulnerabilities CVE-2026-14456, CVE-2026-14457, CVE-2026-18798, CVE-28387, CVE-28388, CVE-28389, CVE-28390, CVE-31789, CVE-31790, CVE-2026-54874, CVE-2026-63072, CVE-2026-63073, CVE-2026-63074, CVE-2026-63075, CVE-2026-63076, and CVE-2026-75803.
The OpenVINO and TensorRT AI² interfaces now provide an SBOM in the form of *.spdx.json files.
The HALCON/.NET interface now provides an SBOM in the form of *.spdx.json files.
For .NET Framework, these reside under bin/dotnet20 and bin/dotnet35.
For .NET Core 3.1 and later, the SBOM is included in the NuGet package under the _manifest directory.
The third-party library libxml has been updated to version 2.13.9.
The third-party library libxslt has been updated to version 1.1.45 to provide a fix for CVE-2025-7424.
Internal security checks identified the following security vulnerabilities in the third-party CUDA SDK:
CVE-2024-0110, CVE-2025-23275, CVE-2025-23308, CVE-2025-23339, CVE-2025-23247, CVE-2024-0111, CVE-2025-23338, CVE-2025-23272, CVE-2025-23274, CVE-2025-23273, CVE-2024-0125, CVE-2024-0124, CVE-2024-0109, CVE-2024-0123, CVE-2024-53878, CVE-2024-53879, CVE-2024-53870, CVE-2024-53871, CVE-2024-53872, CVE-2024-53873, CVE-2024-53874, CVE-2024-53875, CVE-2024-53876, CVE-2024-53877, CVE-2025-23271, CVE-2025-23340, CVE-2025-23248, CVE-2025-23255, CVE-2025-23346.
These vulnerabilities concern components of the CUDA SDK that are neither shipped nor used by HALCON. Exploitation in the context of HALCON is therefore not possible and the CVEs are not applicable for our product.
Internal security checks identified the security vulnerability CVE-2023-6228 (third-party library libtiff). This vulnerability affects the tiffcp utility, which is not distributed as part of HALCON.
Internal security checks identified the security vulnerability CVE-2023-6277 (third-party library libtiff). This is not considered a security vulnerability by the libtiff maintainers, as there is no proven application crash.
Internal security checks identified the security vulnerability CVE-2023-52356 (third-party library libtiff). This vulnerability does not apply to HALCON, as the vulnerable APIs are not used in HALCON.
Exploitation in the context of HALCON is therefore not possible and the CVEs are not applicable for our product.
The third-party library OpenSSL has been updated to version 3.0.18 to provide fixes for CVE-2025-9230 and CVE-2025-9232.
The third-party library cJSON has been updated to version 1.7.19.
Internal security checks identified the following security vulnerabilities in the third-party CUDA SDK:
CVE-2025-33228, CVE-2025-33230, CVE-2025-33229, CVE-2025-33231.
These vulnerabilities concern the NVIDIA Nsight System for Windows and Linux, which are neither shipped nor used by HALCON. Exploitation in the context of HALCON is therefore not possible and the CVEs are not applicable for our product.
The third-party library libxml2 has been updated to version 2.15.2. This fixes the security vulnerabilities CVE-2025-6021, CVE-2025-10911, CVE-2025-7425, CVE-2025-6170.
The third-party library OpenSSL has been updated to version 3.5.5 to provide fixes for CVE-2025-15467, CVE-2025-68160, CVE-2025-69418, CVE-2025-69419, CVE-2025-69420, CVE-2025-69421, CVE-2026-22795, and CVE-2026-22796.
The third-party library zlib has been updated to version 1.3.2 to provide a fix for CVE-2026-27171.
Internal security checks identified the security vulnerability CVE-2026-28500 (third-party ONNX). The vulnerability concerns the Python function onnx.hub.load(), which is neither shipped nor used by HALCON. Therefore, exploitation in the context of HALCON is not possible, and the CVE is not applicable to our product.
The third-party library libpng has been updated to version 1.6.57. This fixes the security vulnerabilities CVE-2025-64505, CVE-2025-64506, CVE-2025-64720, CVE-2025-65018, CVE-2025-66293, CVE-2025-22801, CVE-2026-22695, CVE-2026-25646, CVE-2026-33416, CVE-2026-33636, and CVE-2026-34757.
A Software Bill of Materials (SBOM) in the System Package Data Exchange (SPDX) format has been added next to many binary files with the same name, appended with “.spdx.json”. This includes the MVTec HALCON Image Processing Library and all language interfaces (except HALCON/.NET Core).
The Installation Guide chapter about cybersecurity has been extended with information regarding the secure handling of read operators and system_call.
The release notes now contain an additional section for all issues that are relevant for cybersecurity.
Internal security checks identified the security vulnerability CVE-2013-6648 (third-party Skia). The corresponding function is not used in the production code path. Exploitation in the context of HALCON is therefore not possible and the CVE is not applicable to our product.
Internal security checks identified the security vulnerability CVE-2025-23247 (thirdparty CUDA SDK, cublas). The vulnerability concerns the binary tool cuobjdump, which is shipped alongside the CUDA SDK. HALCON does neither ship nor use this binary. Exploitation in the context of HALCON is therefore not possible and the CVE is for our product not applicable.
The third-party library libtiff has been updated to version 4.7.2. This fixes the security vulnerabilities CVE-2025-8961, CVE-2025-9165, and CVE-2026-4775.
The third-party library thrift has been updated to version 0.23.0. This fixes the security vulnerabilities CVE-2026-41636, CVE-2026-41604, CVE-2025-48431, CVE-2026-41602, CVE-2026-41605, CVE-2026-41603, CVE-2026-41607, CVE-2026-41606, CVE-2026-43870, CVE-2026-43869, CVE-2026-43868.
The third-party library libxml2 has been updated to version 2.15.3. This fixes the security vulnerabilities CVE-2026-6732, CVE-2025-6021, CVE-2025-10911, CVE-2025-7425, CVE-2025-6170.
The third-party library OpenSSL has been updated to version 3.5.7. This fixes the security vulnerabilities CVE-2026-2673, CVE-2026-28387, CVE-2026-28388, CVE-2026-28389, CVE-2026-28390, CVE-2026-31789, CVE-2026-31790, CVE-2025-15467, CVE-2025-68160, CVE-2025-69418, CVE-2025-69419, CVE-2025-69420, CVE-2025-69421, CVE-2026-22795, CVE-2026-22796, CVE-2026-34182, CVE-2026-45447, CVE-2026-42764, CVE-2026-34183, CVE-2026-34180, CVE-2026-7383, CVE-2026-45445, CVE-2026-9076, CVE-2026-34181, CVE-2026-42766, CVE-2026-42767, CVE-2026-42769, CVE-2026-45446, CVE-2026-42768, and CVE-2026-42770.
HDevelop now sends the cross-product UUID, fetched from somctl, when checking for updates and the like.
In HDevelop, accessing a dictionary key after a tuple index operation (such as MyDicts[2].my_key) for either reading or writing did not work. This problem has been fixed. Expressions of such a form are now valid.
We have discontinued the extended telemetry program. Limited telemetry functionality remains in place, as described in our Privacy Policy, and continues to be processed in accordance with applicable data protection requirements.
The HDevelop code export incorrectly exported the 'rep_elem' expression. This problem has been fixed.
Using the Operator Window with incomplete input could crash HDevelop. This problem has been fixed.
The link to the tutorials of the MVTec Academy in the HDevelop start dialog did not work as expected and linked to wrong targets. This problem has been fixed.
HDevelop did not detect an update. This problem has been fixed.
Dragging files like 3D models into HDevelop could create a line of code that led to parser errors. This problem has been fixed.
In some cases, HDevelop crashed when selecting "Discard" in the "New Program" dialog. This problem has been fixed.
HDevelop could crash or hang during startup when using an HDevelop.ini file with content from older versions. This problem has been fixed.
When opening a program editor and then opening and closing a modal dialog (such as the About dialog) using the ‘X’ button, copy, cut, and paste stopped working in the editor. This problem has been fixed.
The procedures dev_display_dl_interactive_confusion_matrix and draw_deep_counting_templates displayed the text on buttons using the same color as the background depending on the set drawing mode. This problem has been fixed.
To show how to use the operator gen_image_warp_map, the new example program hdevelop/Transformations/2D-Transformations/gen_image_warp_map.hdev has been added.
When multiple TensorRT AI² devices were available, the example hdevelop/Deep-Learning/AnomalyDetection/dl_anomaly_detection_global_context_inference.hdev ran into an error. This problem has been fixed, the example now consistently selects the first device for inference.
The HDevelop example hdevelop/ImageSource/image_source_information.hdev has been improved to gather and save all parameter information to file.
The visualization in the HDevelop example program hdevelop/3D-Matching/Deep-3D-Matching/deep_3d_matching_workflow.hdev did not work correctly for cameras of different sizes, or when only one camera was used. This problem has been fixed.
The SBOMs for HDevEngine for the aarch64 and armv7 architectures were missing. This problem has been fixed.
abs_diff_image was slower when NEON optimizations have been enabled using set_system ('neon_enable', 'true') which is the default.
Now, the faster code is used, resulting in a speedup compared to the default by up to 20%.
affine_trans_image, affine_trans_image_size, zoom_image_factor (if ScaleWidth and ScaleHeight are not exactly 0.5), zoom_image_size (if the output image size is not half the input image size), rotate_image (for angles not multiples of 90 degrees), and crop_rectangle2 with AlignToAxis set to 'true' with system parameter 'int_zooming' set to 'true' for interpolation 'bilinear', 'constant', and 'weighted' for images of type 'byte', 'uint2', and 'real', and 'bicubic' for images of type 'real', have been accelerated on Arm 64-bit processors that support the NEON instruction set. A speedup of up to 35% can be expected for images of type 'byte', up to 40% for images of type 'uint2', and up to 55% for images of type 'real'.
The same holds for gen_gauss_pyramid using interpolation 'weighted' for images of type 'byte' and 'constant' and 'weighted' for images of type 'uint2' and 'real', when the Scale is not exactly 0.5.
affine_trans_region with system parameter 'int_zooming' set to 'true' is now up to 35% faster for interpolation 'constant' on Arm 64-bit processors that support the NEON instruction set.
The mean_image filter is now faster by up to 5% for int2 and int4 images on Linux and up to 8% on Windows platforms.
affine_trans_region has been additionally parallelized on internal data level. Previously, it was only parallelized on tuple level. Now, affine_trans_region can be up to twice as fast with two threads, four times faster with four threads, and five times faster with six threads. The overall performance gain depends on the image size, the actual transformation used, and the interpolation method selected.
mean_image for byte images and masks between 7x7 and 15x15 is now faster by up to 35% on AVX2-capable processors.
emphasize for byte images and masks between 7x7 and 15x15 is now faster by up to 25% on AVX2-capable processors.
highpass_image for byte images and masks between 7x7 and 15x15 is now faster by up to 25% on AVX2-capable processors.
projective_trans_image, projective_trans_image_size, and projective_trans_region with system parameter 'int_zooming' set to 'true' for interpolation 'bilinear' on processors that support the Arm NEON instruction set are now up to 45% faster for images of type 'byte'. For interpolation 'nearest_neighbor', a speedup of up to 35% can be expected.
gen_bundle_adjusted_mosaic and gen_projective_mosaic are now up to 45% faster on processors that support the Arm NEON instruction set for images of type 'byte'.
HALCON 25.11 for x86 on Linux and Windows has been compiled with the new Intel clang-based compiler. As a result, many operators are slightly faster than in 25.05, while some are slightly slower. For Linux aarch64, HALCON has been compiled with clang instead of gcc, which also improves overall performance.
fread_bytes and fwrite_bytes are now faster by up to 1500% when reading or writing larger amounts of data.
guided_filter is now up to 115% faster for images of type byte and up to 30% faster for images of type uint2. Due to a different implementation, the results may slightly differ when compared to previous versions. To receive identical results, convert the input images to real.
linear_trans_color has been optimized and is now faster by up to 2000% when parallelized.
zoom_image_size and zoom_image_factor are now faster for images of type byte by up to 180% for x86/x64-capable CPUs if the input image is zoomed up to twice its dimensions and Interpolation is set to 'bilinear', 'constant', or 'weighted'.
fit_primitives_object_model_3d has been extended with new GenParamNames to specify parameters that are assumed to be known and are not changed during the fitting process, namely 'cylinder_axis_dir', 'cylinder_axis_point', 'cylinder_radius', 'plane_normal', 'plane_point', 'sphere_center', 'sphere_radius'.
fit_primitives_object_model_3d is now faster by up to 20% in case that several types of primitives are fitted.
union_object_model_3d has been extended with a new Method 'points_surface_single_viewpoint', which checks mappings for consistency before returning their combination.
Note that this change affects the compatibility. Read more.sample_object_model_3d with Method set to 'xyz_mapping' or 'xyz_mapping_compute_normals' has been improved so that a dense XYZ mapping is available in output for further processing.
Note that this change affects the compatibility. Read more.HALCON has been extended with an operator find_symmetries_object_model_3d that can analyze a 3D object model and determine its symmetric properties.
The new HDevelop example program hdevelop/3D-Matching/Deep-3D-Matching/find_symmetries_object_model_3d.hdev shows how to use the operator to find symmetries and visualize them.
HALCON has been extended with an operator intersect_line_plane_object_model_3d that calculates the intersection point of a line and a plane.
The new HDevelop example program hdevelop/3D-Object-Model/Transformations/intersect_line_plane_object_model_3d.hdev shows how to use this operator.
The content of the result handles of deformable surface matching, i.e., of semantic type 'deformable_surface_matching_result', was not visible in the handle inspect windows. This has been improved. The content of such handles can now be inspected in handle inspect windows.
surface_normals_object_model_3d has been extended with a GenParamName 'xyz_mapping_mask_size' to adjust the width and height of the used neighborhood for speed.
smooth_object_model_3d has accordingly been extended with a GenParamName 'xyz_mapping_compute_normals_mask_size'.
The example hdevelop/3D-Matching/Deep-3D-Matching/deep_3d_matching_training_workflow.hdev now uses advanced object detection. The model match_bitholder.dm3 used by deep_3d_matching_workflow.hdev has been adapted accordingly. The automatic augmentation of (typically synthetically generated) training samples for detector models retrieved from a Deep 3D Matching model has been adapted.
find_surface_model now supports the new parameter 'scene_sampling_mode', which can be set to 'fast' (previous and default behavior) or the new mode 'voxel'. Voxel-based sampling is faster by up to 105% compared to the 'fast' mode, allowing to speed up surface-based matching especially for scenes with a high point density.
Deep 3D Matching now supports the new parameter 'sort_single_views_last', which allows sorting instances seen by only a single camera last in the result list, as they are potentially less reliable and accurate than instances seen by more than one camera.
Note that this change affects the compatibility. Read more.The deep learning procedures dev_display_dl_data and dev_display_dl_data_tiled have been extended for training samples of a 3D pose estimation model retrieved from a Deep 3D Matching model. Now, during the training, result poses on some random samples are displayed in addition to the training curves.
The procedures create_scene_engine_run_params, set_scene_engine_run_param, and check_scene_engine_run_param have been improved regarding parameter checks and error messages.
train_dl_model_batch has been extended with an automatic augmentation of (typically synthetically generated) training samples for detector models retrieved from a Deep 3D Matching model.
The deep learning procedures create_dl_train_param and preprocess_dl_model_images have been adapted so that augmentation in augment_dl_samples is disabled and no normalization is performed.
Note that these changes do not affect standard detector models, which are not used in Deep 3D Matching.
open_scene_engine has been extended with the generic parameter 'debug_log_file'. When setting a string specifying filename and location of a debug file, the opened Scene Engine instance creates the corresponding file and writes debug information to it.
Handles of type class_lut and color_trans_lut can now be serialized and deserialized using the operators serialize_handle and deserialize_handle.
In the Handle Inspect window, now some information about handles of type class_lut is visible. This information can be queried for debugging purposes using the operators get_handle_param and get_handle_tuple.
HALCON's advanced object detection has been extended to support class specific minimum confidence thresholds at inference.
For advanced object detection during inference, proposals with confidence scores below min_confidence are discarded. Models now support class-specific values of min_confidence, allowing the confidence threshold to be configured independently for each class.
Suggestions for minimum confidence thresholds can be calculated from the validation split of the dataset with the new procedure determine_dl_model_detection_confidence_thresholds.
The HDevelop examples for advanced object detection now demonstrate the necessary configuration to use class-specific min_confidence values. The HDevelop examples for advanced object detection also offer the possibility to visualize precision and recall values for different minimum confidence values.
HALCON's deep-learning-based advanced object detection has been extended with oriented object-detection to locate instances of oriented objects in an image.
Advanced object detection for oriented instances comes with three pretrained models:
This new approach is significantly faster than the previous one on both GPU and CPU runtimes as well as on AI² interfaces. The new anchor-free network architecture is easier to use because no anchor parameters have to be tuned.
Further, the new models shows superior precision and recall on testing datasets due to an improved network architecture and a pretraining on a large-scale dataset.
This update replaces the previous object detection for oriented bounding box prediction. Now, creating object detection models with create_dl_model_detection is no longer needed for rectangle2 instances.
The HDevelop example hdevelop/Deep-Learning/Detection/dl_advanced_object_detection_with_orientation_workflow.hdev has been added to demonstrate how to use advanced object detection for oriented instances in HALCON.
The Deep OCR detection component now has an improved orientation estimation, resulting in more accurate bounding boxes.
query_available_dl_devices reported unsupported TensorRT GPU devices with compute capability lower than 7.5 for the AI² TensorRT plugins 25.11.1 and 25.11.2. This problem has been fixed with the AI² TensorRT plugin 25.11.3 and newer.
When optimizing deep learning models for TensorRT devices, now the hardware_compatibility_level parameter is available.
The cross entropy loss layer now supports soft targets in the form of class probability vectors.
transform_dl_sample_batch is now parallelized on batch item level.
HALCON has been extended with a new AI Accelerator Interface (AI²) featuring the ONNX Runtime.
In this release, the plugin enables optimized CPU inference through the ONNX Runtime CPU Execution Provider.
The AI² can be installed using the MVTec Software Manager and is supported on the following platforms:
x64-win64, x64-linux, aarch64-linux
The plugin is documented online at https://docs.mvtec.com. The documentation features a description of the new plugin, including installation, configuration, and usage instructions.
A new HDevelop example, hdevelop/Deep-Learning/AI2Interfaces/run_deep_ocr_on_onnxruntime_cpu_ep.hdev, demonstrates how to perform deep learning inference, such as Deep OCR, using the ONNX Runtime AI² CPU Execution Provider.
These operators have been extended to support the new AI² interface:
query_available_dl_devices, optimize_dl_model_for_inference, get_dl_device_param, set_dl_device_param
Enhanced Data Augmentation introduces a more flexible, operator-based augmentation system for HALCON deep-learning workflows. You can build configurable augmentation pipelines and preview the transformations, which makes it easier to iterate quickly and integrate augmentation cleanly into existing DL training setups. The initial scope targets Object Detection and Instance Segmentation, using modern techniques such as color (HSV) variation, flips, crops, perspective transforms, blur, and combinations to improve robustness under real-world variability, such as lighting changes, viewpoint shifts, and occlusions, often reducing the amount of training data needed. Since this moves from procedure-based setups to operator-based pipelines, existing example workflows remain a key compatibility point and custom workflows may need small adjustments.
The example hdevelop/Deep-Learning/Augmentation/dl_transform_pipeline.hdev has been added to illustrate how to visualize the output of the new augmentation pipeline. The example hdevelop/Deep-Learning/Detection/dl_advanced_object_detection_workflow.hdev demonstrates how the new workflow can be applied in practice and includes a use case showing the new augmentation in action.
To create the new transform pipelines, this operator has been added:
To define the single transform steps, seven operators are now available:
To get and set transform parameters or read and write the pipeline, these operators have been introduced:
The transform pipeline can be executed with:
Additionally, two new deep learning procedures are now available. One to get the transformed DLSamples and one to display and inspect the transform pipeline:
HALCON has been extended with deep-learning-based advanced object detection to locate object instances in an image.
Advanced object detection comes with three models, pretrained for object detection.
This new approach is significantly faster than the previous one on both GPU and CPU runtimes as well as on AI² interfaces.
The new anchor-free network architecture is easier to use because no anchor parameters have to be tuned.
Further, the new models showed superior precision and recall on testing datasets due to an improved network architecture and pretraining.
This update replaces the previous object detection for axis-aligned bounding box prediction.
Now, creating object detection models with create_dl_model_detection is no longer needed for rectangle1 instances.
The HDevelop example hdevelop/Deep-Learning/Detection/dl_advanced_object_detection_workflow.hdev has been added to demonstrate how to use advanced object detection in HALCON.
The HDevelop example series hdevelop/Deep-Learning/Detection/detect_pills_deep_learning_*.hdev has been extended to advanced object detection. The example covering the old object detection, hdevelop/Deep-Learning/Detection/dl_object_detection_workflow.hdev, has been removed.
The reference manual entry “Deep Learning” > “Advanced Object Detection” has been added, providing instructions on how to configure and use the new object detection models.
For more information on specific usage, please refer to the documentation for the respective operators:
set_dl_model_layer_param now supports a new parameter 'coefficients' to set the coefficients of an elementwise layer.
The pretrained Deep OCR recognition models have been updated and now offer improved compatibility with INT8 calibration. Furthermore, it is now possible to recognize whitespace between words.
Deep learning models of type 'segmentation' now support the parameters 'tiling' and 'tiling_overlap' to enable automatic tiling of large images.
The Deep OCR recognition component now returns more reliable confidences for character candidates.
The pretrained MobileNetV4 classification models have been updated and now offer improved compatibility with INT8 calibration.
create_dl_layer_loss_ctc now accepts 'loss_weight' as an optional parameter. Its scalar value can be used to weigh its contribution to the overall loss.
The HDevelop example hdevelop/Deep-Learning/Detection/dl_detection_inference_yolo.hdev has been adapted. The model in the example does not need a special treatment in case it is executed on an AI² device anymore.
set_dl_model_param now supports a new parameter 'fuse_conv_bn' to fuse pairs of convolution and batch normalization layers. This fusion can improve the runtime performance and memory consumption of the model during inference.
read_dl_model has been extended such that the ONNX MatMul operation can now be read as a MatMul layer in HALCON.
For some deep learning functionality, now the “cuDNN FrontEnd” (FE) API by NVIDIA is used. This implies that a computational graph is compiled at runtime (JIT) for specific operations and given GPU hardware. Due to this online optimization process, the device initialization of a given model using such a layer or functionality takes much longer than before.
create_dl_layer_elementwise now supports unidirectional broadcasting for the elementwise operation 'sum'. Further, for all commutative elementwise operations, unidirectional broadcasting is supported for any order of input layers.
HALCON has been expanded to include Continual Learning for all deep learning classification models. This function allows models to be incrementally updated with new data while retaining previously learned knowledge.
In addition to the extension of the deep learning operators and procedures with the new Continual Learning functionality, the operators init_dl_continual_learning and extend_dl_continual_learning have been introduced.
The HDevelop examples hdevelop/Deep-Learning/Classification/continual_learning_for_classification.hdev and hdevelop/Deep-Learning/Classification/continual_learning_for_classification_on_edge_device.hdev have been added to demonstrate how to configure and utilize a classification deep learning model with this feature.
The reference manual entry "Deep Learning" > "Continual Learning" has been added, providing instructions on how to configure and use a classification model with Continual Learning functionality.
set_dl_model_param now supports a new parameter 'max_gradient_norm' to set a limit for the gradient norm of the weights and biases of the model during training. This is also known as gradient clipping.
create_dl_layer_permutation and create_dl_layer_slice have been extended to support a new behavior, which allows specifying a permutation or axes in 'nchw' format. This behavior can be enabled by setting the GenParam 'axes_format' to 'nchw'.
The NVIDIA CUDA libraries used by HALCON's Deep Learning Framework have been updated to CUDA 12.8 and cuDNN 9.10 on x64-linux and x64-win64 platforms. With this release, the NVIDIA Blackwell architecture is supported.
On aarch64 systems, Jetpack 6.2 is required.
The TensorRT AI² interface has been updated to use TensorRT 10.12.0 to support the Blackwell architecture as well.
Note that this change affects the compatibility. Read more.The parameter 'ai_accelerator_interface' of the operators query_available_dl_devices and get_dl_device_param has been renamed to 'ai2'.
set_dl_model_param now supports a new parameter 'weight_decay' to enable decoupled weight decay, a popular regularization method for model training.
create_dl_layer_loss_ctc now supports a new parameter 'zero_infinity' to suppress batch items with infinite loss, which can stabilize the training. Furthermore, the CTC loss layer now provides better determinism and better balancing of samples with different sequence lengths. This improves the training of Deep OCR recognition models.
Note that this change affects the compatibility. Read more.create_dl_layer_convolution and create_dl_layer_activation now support Tanh activation.
create_dl_layer_concat now allows a single input layer.
read_dl_model has been extended such that the ONNX Tanh operation can now be read as an activation layer in HALCON.
HALCON has been extended with pretrained deep learning classification models from the MobileNetV4 family. Depending on runtime and accuracy requirements, the user can choose between pretrained_dl_model_mobilenet_v4_small.hdl, pretrained_dl_model_mobilenet_v4_medium.hdl, and pretrained_dl_model_mobilenet_v4_large.hdl. All models are also available as backbones for object detection.
The DL framework now contains the new operator create_dl_layer_gather, which creates a Gather layer within a custom DL model. The layer gathers values from the input data along a specified axis according to the given indices.
read_dl_model has been extended such that the ONNX Gather operation can now be read as a Gather layer in HALCON as long as the indices have a rank smaller than 2.
create_dl_layer_input and set_dl_model_layer_param now support a new generic parameter 'keep_batch_size_const', which can be set to 'true' for input layers of type 'constant'. In this case, the batch size of this layer is not affected by changes of the model batch size.
HALCON now supports reading and running inference of YOLO v8/v9/v11 Deep Learning object detection models (axis-aligned or oriented bounding boxes). YOLO models are known for their real-time speed and high accuracy, making them ideal for applications requiring quick and reliable object detections. To use this feature, a trained YOLO ONNX model is required.
The existing example hdevelop/Deep-Learing/Detection/dl_detection_inference_yolo.hdev has been extended, which guides you through the workflow.
The GenParam scaling parameter of gen_dl_model_heatmap has been extended to include the option 'scale_without_relu'. In addition, the default value for this parameter is now set to 'scale_without_relu'. The scaling parameter is now also applicable when HeatmapMethod is set to 'guided_grad_cam'.
create_dl_layer_activation has been extended with the activation modes 'sin' and 'cos'.
read_dl_model has been extended such that the ONNX Sin and Cos operations can now be read as activation layer in HALCON.
create_dl_layer_activation now supports 31 more activation modes:
'asin', 'acos', 'atan', 'abs', 'ceil', 'celu', 'clip', 'cos', 'cosh', 'elu', 'erf', 'exp', 'floor', 'gelu', 'hard_sigmoid', 'hard_swish', 'log', 'mish', 'neg', 'pow', 'reciprocal', 'round', 'sin', 'sinh', 'softplus', 'softsign', 'sqrt', 'swish', 'tan', 'tanh', 'thresholded_relu'.
The corresponding ONNX operations are supported in the operator read_dl_model.
fit_dl_out_of_distribution has been updated to support datasets containing fewer than 15 training samples per class, as long as validation samples for that class are available.
The forward pass on CPUs of the elementwise layer has been accelerated by parallelizing the elementwise operations 'sum' with broadcasting, 'minimum', 'maximum', and 'division'.
The pretrained Deep OCR recognition model has been improved and extended. The runtime and memory consumption have been significantly reduced, while the accuracy has been slightly improved.
Additionally, a new compact recognition model has been added. This compact model is optimized for inference on low-power devices, prioritizing computational efficiency over accuracy. Both models can be retrained using Deep OCR recognition training.
set_deep_ocr_param has been extended with the option to specify ‘compact’ or ‘default’ for the parameter ‘recognition_model’.
The following pretrained model has been updated: dl/pretrained_deep_ocr_recognition.hdl
The following pretrained model has been added: dl/pretrained_deep_ocr_recognition_compact.hdl
HALCON has been enhanced with Continual Learning for all deep learning classification models. This function allows models to be incrementally updated with new data while retaining previously learned knowledge.
In addition to the extension of the deep learning operators and procedures with the new Continual Learning functionality, the operators init_dl_continual_learning and extend_dl_continual_learning have been introduced.
The HDevelop examples hdevelop/Deep-Learning/Classification/continual_learning_for_classification.hdev and hdevelop/Deep-Learning/Classification/continual_learning_for_classification_on_edge_device.hdev have been added to demonstrate how to configure and utilize a classification deep learning model with this feature.
The reference manual entry “Deep Learning” > “Continual Learning” has been added, providing instructions on how to configure and use a classification model with Continual Learning functionality.
HALCON now uses OpenJPEG to read JPEG 2000 images instead of libjasper, as previous versions of HALCON did.
Note that this change affects the compatibility. Read more.edges_sub_pix and edges_color_sub_pix were improved by adding a suppression of small, spurious edge contours, for example, to facilitate postprocessing. The suppression was added to all subpixel edge operators except 'sobel_fast'.
HALCON now provides ways to smooth images in channel direction as well as to compute derivatives in channel direction. This is especially helpful for multispectral and hyperspectral images. The new operator convol_channels can be used to apply a linear convolution filter in channel direction. The new operator gen_savitzky_golay_filter can be used to create filter parameters for such a convolution.
The new HDevelop example hdevelop/Filters/Color/smooth_hyperspectral_images.hdev shows how to use the new operators to smooth the data and to compute derivatives in channel direction.
HALCON now provides a way to use canonical variates analysis on the channels of an image to reduce the dimensionality while keeping target classes separable. This is especially helpful in the analysis of multi- or hyperspectral images. The new operator gen_canonical_variates_trans uses the canonical variates analysis to compute a linear transformation, which can be applied using linear_trans_color.
The new HDevelop example hdevelop/Filters/Color/analyze_pills_hyperspectral_canonical_variates.hdev shows how to use this functionality to reduce hyperspectral images to RGB images.
The Postscript window type is no longer supported and has been removed.
Note that this change affects the compatibility. Read more.HALCON has been extended with the operator open_runtime_window. It opens a new graphics window and makes open_window more extensible by using a dictionary as input. HDevelopEVO extends the functionality of the operator to arrange graphic windows in the IDE. Hence, open_runtime_window serves as the common ground for managing graphics windows in HALCON applications, which facilitates re-use and export of HALCON scripts written with HDevelopEVO.
The Data Matrix reader has been extended with AI-based candidate detection. This enables a more robust detection of codes in challenging environments, such as low contrast or nearby clutter.
To activate this feature, the 'detection_method' parameter must be set to 'deep_learning' when creating the Data Matrix reader handle using create_data_code_2d_model. Otherwise, the default 'classic' rule-based method will be used.
The pretrained model pretrained_data_matrix_detection.hdl included with HALCON uses tiling by default and can be further optimized using AI² interfaces.
The minimum confidence for Data Matrix code candidates can be set with the new parameter 'detection_min_confidence' or trained with the existing training mechanism in find_data_code_2d.
A new example has been added showing the workflow with AI-based Data Matrix detection:
hdevelop/Identification/Data-Code/ecc200_deep_learning_detection.hdev
For this example, the new images images/datacode/ecc200/ecc200_dongle_01.png to images/datacode/ecc200/ecc200_dongle_08.png have been added.
The QR code reader is now more robust with branded codes.
The 'robust' decoding algorithm used in ISO/IEC 15415 and ISO/IEC 29158 print quality inspection of Data Matrix codes has been improved. Especially the boundary of the symbol is now detected more robustly to achieve more accurate grades.
This might lead to slightly different grades compared to earlier versions. However, the overall grading is more robust with these improvements.
Further, the determination of the position of the non-data module centers of Data Matrix codes in ISO/IEC 15415 and ISO/IEC 29158 print quality inspection has been adapted to be in line with the new 2024 revision of ISO/IEC 16022.
HALCON's Data Matrix reader is now more robust on deformed codes, especially on codes with multiple data regions and dot printed codes.
The reference decode algorithm of the Data Matrix ECC 200 usable by setting the set_data_code_2d_param parameters 'quality_isoiec15415_decode_algorithm' or 'quality_isoiec29158_decode_algorithm' to 'reference' has been updated to follow the newest revision (see ISO/IEC 16022:2024). The changes made can lead to different grading results.
The Data Matrix, QR Code, and DotCode reader now require less memory. The reduction depends on the parameter setting and is highest with standard_recognition and without parallelization. The memory requirements for symbol_search=rudimental have not been changed.
The performance of these readers slightly improves in these situations as well. In best cases, the QR Code reader is 10 percent faster.
HALCON's DataMatrix reader is now more robust when multiple codes are positioned very close to each other in an image.
HALCON's DataMatrix reader is now more robust on codes with perspective distortion. The parameter 'alternating_pattern_tolerance' toggles this improved functionality. If this parameter is set to 'high', the improved functionality is used. For the settings 'low' and 'medium', the behavior is not changed.
The training mode of the operator find_data_code_2d has been improved. Now, training calls require less memory. This improvement does not change the results of the training.
HALCON's Data Matrix reader is now more robust on deformed codes. The parameter 'deformation_tolerance' has been added to toggle this new functionality. With the default mode 'maximum_recognition', this parameter is set to 'curved'. This parameter is trainable.
To ease debugging of this new functionality, get_data_code_2d_objects has been extended with 'candidate_xld_deformed' to return a more accurate estimation of the symbol boundary. get_data_code_2d_results has been extended with 'deformation_objects' to return a dict with a rectified image of the code and respective module regions.
The example hdevelop/Identification/Data-Code/2d_data_codes_default_settings.hdev has been extended with two new images: images/datacode/ecc200/ecc200_deformed_01.png and images/datacode/ecc200/ecc200_deformed_02.png.
HALCON's 2D code readers can now heuristically determine the string encoding of an encoded message. set_data_code_2d_param has been extended with the new value 'auto' for the parameter 'string_encoding'. This is now the default for this parameter.
set_data_code_2d_param has also been extended with the new value 'gb2312' for the parameter 'string_encoding' to support GB2312 encoded messages.
The parameter 'contrast_min' has been removed from the (Micro) QR Code reader as its influence on the decoding performance was negligible after the improvements of the last releases. The training of 'contrast' has been removed for these readers as well. It is not possible to query 'contrast' with get_data_code_2d_results for the (Micro) QR Code reader anymore.
Note that this change affects the compatibility. Read more.HALCON now implements the newest standard ISO/IEC 15415:2024 for print quality inspection of two-dimensional barcodes by calling the operator get_data_code_2d_results with the parameter 'ResultNames' set to 'quality_isoiec15415'.
The new version of the print quality inspection standard has an impact on the Data Matrix ECC 200, QR Code, Micro QR Code, Aztec Code, and PDF417 code readers and the related operators, parameters, and results:
Furthermore, the following standard examples were adapted with respect to ISO/IEC 15415:2024:
HALCON's QR Code reader is now even more robust on deformed codes when setting the parameter 'deformation_tolerance' to 'curved'. By using a more powerful method to find code candidates in the image, the detection rate has been increased significantly, especially for challenging scenarios like images with strongly distorted codes.
Further, the detection rate has been significantly increased for standard and enhanced recognition without using the parameter 'deformation_tolerance', while simultaneously reducing the run time.
HALCON now implements the newest standard ISO/IEC 29158:2025 for assessing the symbol quality of direct marked parts (DPM). The assessment can be executed by calling the operator get_data_code_2d_results with the parameter 'ResultNames' set to 'quality_isoiec29158'.
The new version and the changes made affect the assessment of the Data Matrix ECC 200, QR Code, Micro QR Code, and Aztec Code. The affected operators, procedures, parameters, and results are:
Furthermore, the following standard examples have been adapted and renamed with respect to ISO/IEC 29158:2025:
Similarly, the images datacode/ecc200/ecc200_quality_aimdpm_[01-08].png have been renamed to ecc200_quality_isoiec29158_[01-08].png.
Note that this change affects the compatibility. Read more.The behavior of the 'strict_quiet_zone' check, which can be enabled via set_data_code_2d_param, changed due to the adaptations made regarding the print quality inspection standards ISO/IEC 15415 and ISO/IEC 29158. This affects all 2D code types supporting this parameter.
The bar code reader has been improved for GS1 DataBar Omnidir, GS1 DataBar Truncated, GS1 DataBar Limited, and GS1 DataBar Expanded codes with composite component. The composite part can now be detected more robustly.
find_bar_code in conjunction with the parameter 'composite_code' set to 'CC-A/B' allowed to decode composite symbols of the GS1 DataBar family. Now, the barcode types EAN-8, EAN-13, UPC-A, UPC-E, and GS1-128 support composite structures of type A and B as well.
set_bar_code_param has been extended with the generic parameter 'strict_ean13_upc_a'. This parameter defines the distinction of barcodes of type EAN-13 and barcodes of type UPC-A.
Previously, UPC-A barcodes had been handled as a subset of EAN-13 barcodes. Most notably their decoded data string did consist of 13 digits instead of the expected 12 digits.
Now, 'strict_ean13_upc_a' can be set to either 'false' or 'true'. Setting it to 'false' keeps the previous behavior. Setting it to 'true' enforces a strict distinction between these two types of barcodes as described in ISO/IEC 15420:2009, Table 3.
get_bar_code_object now allows querying 'scanlines_identical' and 'scanlines_identical_plain'. 'scanlines_identical' shows a subset of 'scanlines_valid', defined by the scanlines, which were needed to reach 'min_identical_scanlines' (see set_bar_code_param).
The bar code reader has been improved for Code 128 and GS1-128 codes with irregular element widths. Now, the bar code model parameter 'element_size_variable' influences the decoding of these code types. Setting 'element_size_variable' to 'true' enables find_bar_code to compensate small variations in the element size, which can help to increase the decoding rate. Further, the parameter 'element_size_variable' can now be specifically set and queried for different bar code types by using the operators set_bar_code_param_specific and get_bar_code_param_specific.
For the print quality assessment of direct part marked Data Matrix ECC 200 symbols (ISO/IEC 29158 standard), the new set_data_code_2d_param parameter 'quality_isoiec29158_decode_algorithm' has been introduced. This parameter offers the possibility to switch between a 'robust' and a 'reference' implementation. While the 'robust' mode represents the previous HALCON implementation, the new 'reference' mode is strictly following the reference decode algorithm described in ISO/IEC 16022, including the adaptations of ISO/IEC 29158. The latter is intended for grading with a standard-compliant camera and lighting setup.
The new mode uses the so-called dot connecting algorithm described in Annex C of ISO/IEC 29158 to ensure continuous edges even for dot peen symbols, where the reference decode algorithm would fail otherwise. If the dot connecting algorithm led to the best grades, the user will be notified by a value 'true', available via get_data_code_2d_results and the parameter 'quality_isoiec29158' or 'quality_isoiec29158_values'. The corresponding name of the new entry is called 'Dot connecting algorithm used' and can be queried via 'quality_isoiec29158_labels' as usual.
The following examples have been adapted to use the new 'reference' mode:
Minor additional changes: The "Solution Guide II-C 2D Codes" mentions the new parameter 'Dot connecting algorithm used'. Furthermore, the reference decode algorithm as a whole has been partly reworked. These changes can lead to slight differences in the grading not only for ISO/IEC 29158 but also for ISO/IEC 15415 following the reference decode algorithm (see the parameter 'quality_isoiec15415_decode_algorithm'). The overall runtime of the reference decode algorithm has been improved as well.
The following GS1 Application Identifiers are now supported:
gen_image_warp_map can now be executed on OpenCL devices.
paint_xld is now more accurate if called with multiple contours that contain touching edges.
The runtime of paint_xld is now independent of the setting of 'init_new_image'.
access_channel now allows selecting multiple channels of the input image. Additionally, multiple multi-channel input images can be passed, in which case the same channels are selected from all of them.
HALCON shape-based matching now supports the use of 'max_deformation' during contour pruning.
This enhancement extends the Automatic Contour Optimization workflow to applications where objects show moderate geometric deviations between the template and the training samples. As a result, contour pruning can create robust matching models even when sample images contain manufacturing tolerances, perspective effects, or small object deformations.
To use this feature, set the 'max_deformation' parameter with set_generic_shape_model_param before training. No additional training parameters are required and the existing contour pruning workflow remains unchanged. The contour pruning algorithm now takes the specified deformation tolerance into account when finding the model on the provided samples. This allows stable contour information to be identified more reliably while still removing unstable contour parts.
The result is a contour-pruned model even for slightly deformed objects, reducing manual contour editing effort and extending the applicability of Automatic Contour Optimization to a wider range of applications.
find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, find_aniso_shape_models, and find_generic_shape_model differed in the evaluation of match candidates during tracking for different simd implementations. Now their results deviate less.
The generic shape matching interface now allows restricting all models passed in find_generic_shape_model by a global number of matches. To enable global num_matches, call set_generic_shape_model_param with 'num_matches_global_enable' set to the value 'true'.
HALCON shape-based matching has received a new class of algorithms: Automatic Contour Optimization.
This feature optimizes the model contours during training using the sample-based approach from extended parameter estimation. In addition, 'extended parameter estimation' is renamed to 'sample-based training' to better reflect the functionality of the existing and new algorithms and to allow access through one unified interface.
To query if the current shape model setting and the desired sample-based training methods are compatible, the operator get_generic_shape_model_param has been extended with the parameter 'sample_based_training_compatible'.
The first method for Automatic Contour Optimization is the 'contour pruning' algorithm, which removes unstable contour parts of the model based on the provided samples. To use the new algorithm, set the corresponding value in the 'sample_based_training' parameter (formerly known as 'extended_parameter_estimation') and add the samples using set_generic_shape_model_param. The next train_generic_shape_model call will then apply the contour pruning during training. The result is a model that only keeps the parts of your model that are stable, which can lead to a faster, more robust and more accurate matching model.
To configure the contour pruning algorithm, the new 'contour_pruning_threshold' parameter can be set with a set_generic_shape_model_param call. This value determines which score each part of the model has to reach to still be considered for the optimized model contours.
The new HDevelop example hdevelop/Matching/Shape-Based/train_generic_shape_model_contour_pruning_workflow.hdev shows how to use this feature in HALCON.
HALCON shape-based matching has been extended with Score Visualization for Shape Matching.
The model contour parts with model points contributing to the score within a set score interval can be retrieved. For this search feature, set_generic_shape_model_param has been extended to set the parameter 'score_visualization' to calculate the contributions during find_generic_shape_model.
get_generic_shape_model_result has been extended to retrieve the contours.
To query if a shape model supports the feature, the call get_generic_shape_model_param has been extended with the parameter 'score_visualization_enabled'.
The new example
hdevelop/Matching/Shape-Based/get_generic_shape_model_result_score_visualization_contours.hdev
has been added to show how to use this feature in HALCON.
The default memory allocator on Windows systems is no longer mimalloc, but the system allocator. This is the same behavior as on Linux systems.
Note that this change affects the compatibility. Read more.When adding a message to a message queue that is already full, the operator enqueue_message can now wait and block until space becomes available in the queue. This allows using producer-consumer threads that automatically throttle the producer to the rate of the consumer.
Note that this change affects the compatibility. Read more.tuple_rand(0) will now return an empty tuple instead of throwing an error.
Note that this change affects the compatibility. Read more.When rerunning deep_3d_matching_training_workflow.hdev or resuming an existing training, preprocess_dl_dataset unnecessarily reprocessed the complete dataset instead of reusing existing preprocessed data. This problem has been fixed. Existing preprocessed data is now reused if the dataset and preprocessing configuration have not changed.
read_object_model_3d may have crashed for large PCD files. This problem has been fixed.
dev_display_dl_data did not display the text 'gripping_map'. dev_display_dl_3d_data did not display 'gripping_point_cloud'. These problems have been fixed.
project_object_model_3d in rare cases created artifacts and did not return all expected contours for telecentric camera parameters and 'hidden_surface_removal' set to 'true'. This problem has been fixed.
fit_primitives_object_model_3d and segment_object_model_3d in rare cases may have crashed or returned no result. This problem has been fixed.
connection_object_model_3d sometimes did not raise errors or did not unlock models in error cases and might have crashed in rare cases. This problem has been fixed.
The accuracy of poses detected and refined with Deep 3D Matching, i.e., with the operator apply_deep_matching_3d, has been improved.
When reading and writing OBJ files with read_object_model_3d and write_object_model_3d, point normals were not correctly handled when the file contained triangles or polygons. This problem has been fixed. Triangles and polygons now have the correct per-point attributes. The ability to import pure point cloud data (no lines or faces) with normals is unaffected.
Additionally, read_object_model_3d and write_object_model_3d for OBJ files have been extended to support reading and writing of line data as well as per-vertex texture coordinates.
distance_object_model_3d with 'store_closest_index'='true' and 'distance_to'='triangles' returned an incorrect closest index if ObjectModel3DTo contained triangles and the attribute '&original_indices', but no polygons. This occurred when ObjectModel3DTo was processed with triangulate_object_model_3d 'polygon_triangulation' or created by read_object_model_3d from an .obj file with 'convert_to_triangles'='true'. This problem has been fixed. Now, triangle indices are returned.
preprocess_dl_samples and preprocess_dl_model_3d_data returned faulty normals when XYZ maps contained NaN values. This problem has been fixed. Now, pixels with NaN values are excluded from the normals calculation.
find_surface_model, find_surface_model_image, refine_surface_model_pose, and refine_surface_model_pose_image with 'pose_ref_use_scene_normals'='true' did not differentiate enough between normals pointing in opposite directions in pose refinement. This problem has been fixed.
object_model_3d_to_xyz with Type 'cartesian_faces_no_opengl' in rare cases returned the error 3513 ("Internal error: number of chords too big for num_max"). Further, object_model_3d_to_xyz with Type 'cartesian_faces' was not as accurate as possible for telecentric lenses. These problems have been fixed.
reconstruct_surface_stereo in rare cases crashed or raised a memory error with enabled 'persistence' of the stereo model. This problem has been fixed.
run_scene_engine returned an empty visibility tuple and no masks image in GeneratedData.images when no part objects were visible in a generated scene. This led to errors in subsequent procedure calls of visualize_scene_engine_run and get_data_generation_gt. This problem has been fixed. Now, run_scene_engine returns a tuple of zeros as visibility and an appropriate masks image.
apply_deep_matching_3d did not always return all poses seen from single cameras when setting the parameter 'min_num_views' of a Deep 3D Matching model to 1 and using multiple cameras. This problem has been fixed.
projective_trans_image_size with TransformDomain='true' crashed in rare cases. This problem has been fixed.
binocular_disparity_ms, binocular_distance_ms, and reconstruct_surface_stereo with 'disparity_method' set to 'binocular_ms' in rare cases returned different results if called consecutively. This problem has been fixed.
connection_object_model_3d did not raise an error when more than one Feature was passed or Value was out of bounds. This problem has been fixed.
set_deep_matching_3d_param with GenParamName 'camera_pose N' only worked correctly for poses of type 0. Pose types 1 to 13 did not work correctly. This problem has been fixed.
remove_object_model_3d_attrib and remove_object_model_3d_attrib_mod with attributes 'triangles', 'lines', or 'polygons' did not remove extended attributes attached to these triangles, lines, or polygons. This inconsistent stage could lead to problems in subsequent operators like select_points_object_model_3d or union_object_model_3d. This problem has been fixed. Now, extended attributes attached to triangles, lines, or polygons are removed if 'triangles', 'lines', or 'polygons' are removed, respectively.
find_deformable_surface_model, find_surface_model, find_surface_model_image, and register_object_model_3d_pair in very rare cases leaked memory when their execution was canceled or interrupted. This problem has been fixed.
create_deep_matching_3d with symmetry poses of non-mixed tuple type or poses of types not equal to 0 did not work correctly. This problem has been fixed.
The procedure create_dl_train_param raised an error for models of type '3d_pose_estimation' for empty GenParamName and GenParamValue. In particular, it raised an error for missing generic evaluation parameters even if EvaluationIntervalEpochs was set to 0. This problem has been fixed.
gen_binocular_rectification_map and its internal call in set_stereo_model_image_pairs in rare cases returned very large translations in the rectified poses for telecentric cameras, sometimes resulting in numerical artifacts in a later reconstructed point cloud. This problem has been fixed.
apply_deep_matching_3d returned the error 8406 ("Point cannot be projected") in rare cases. This problem has been fixed.
The procedures check_dl_preprocess_param, preprocess_dl_dataset, and gen_dl_samples did not treat the parameter 'min_visibility' correctly, leading to errors in the detection training in example deep_3d_matching_training_workflow.hdev. This problem has been fixed.
For open_scene_engine, the default timeout of 10 seconds was insufficient on some systems. The default timeout has been increased to 60 seconds.
The procedure set_scene_engine_run_param sometimes threw error 1401 ("Wrong number of values of control parameter 1") in operation 'max'. This problem has been fixed.
prepare_object_model_3d with Purpose set to 'shape_based_matching_3d' in rare cases added an empty shape attribute instead of raising an error.
create_shape_model_3d in rare cases added an empty shape attribute when it raised an error. These problems have been fixed.
When passing a 3D object model with pre-defined vertex normals to the scene engine, those normals were not always used and copied when the physics engine stage was active. This problem has been fixed.
run_scene_engine returned object instance visibilities that contain NaN ("Not a Number") values in some cases. This problem has been fixed. All instance visibilities returned by run_scene_engine are now in the range 0 to 1.
When using run_scene_engine on a system without a suitable GPU, the Scene Engine returned black images and printed errors about missing shaders to the console. This problem has been fixed. Now, the required shaders are part of the installation, and images can be rendered on CPU. However, using a GPU is recommended for performance reasons.
In some cases, the brightness of images rendered with the scene engine could vary when rendering repeatedly with the same scene engine instance. This problem has been fixed. The image brightness is now consistent, even when rendering multiple images.
When running run_scene_engine, in some cases parts of the ground and the distractor objects were rendered too dark. This problem has been fixed.
uncalibrated_photometric_stereo did not check the input parameter ResultType. This problem has been fixed. Now, only ResultType 'all' is supported.
Note that this change affects the compatibility. Read more.triangulate_object_model_3d with 'Method' set to 'greedy' crashed in very rare cases. This problem has been fixed.
find_calib_object crashed in very rare cases. This problem has been fixed.
Passing object detection training results to the dev_display_dl_model_train_info procedure led to an error. This problem has been fixed.
transform_dl_sample_batch for large RangePerspective specified in create_dl_transform_random_geometric sometimes did not transform labels or raised internal errors. This problem has been fixed.
apply_dl_model and train_dl_model_batch led to a crash in 'cpu' devices for image_width=32 on Advanced Object Detection (AOD) models. This problem has been fixed.
After classification models had been optimized for AI² inference with optimize_dl_model_for_inference, re-fitting Out-of-Distribution (OOD) Detection was not possible. This problem has been fixed.
Users can now recalibrate OOD detection on AI²-optimized models to compensate for accuracy changes introduced by the optimization process.
The generic parameters class_weights and max_weight were ignored in the procedure preprocess_dl_dataset for 3d_gripping_point_detection models. This problem has been fixed.
The generic parameters class_weights and max_weight were ignored in the procedure preprocess_dl_dataset for overwrite_files set to auto. This problem has been fixed.
Training a deep learning model with a zoom layer in 'nearest_neighbor' mode on the 'gpu' runtime was unexpectedly slow due to a bug in the backward pass. This problem has been fixed.
For Deep OCR models, the postprocessing parameters detection_min_word_area and detection_tiling_overlap could neither be set via set_deep_ocr_param nor via set_dl_model_param after optimize_dl_model_for_inference had been called. However, these parameters do not affect the underlying model structure and, hence, should be settable. This problem has been fixed.
The procedure calculate_dl_segmentation_class_weights incorrectly returned an error if the parameter IgnoreClassIDs contained IDs that did not belong to DLDataset.class_ids. This problem has been fixed.
Reading ONNX models with read_dl_model failed if the ONNX model contained operations where the axis specification was negative and the input node of the operation had rank < 4. This problem has been fixed.
Reading ONNX models containing elementwise layers with inputs that specify different numbers of explicit dimensions could trigger an error due to a dimensional mismatch. This problem has been fixed.
The HDevelop example hdevelop/Deep-Learning/Detection/dl_detection_inference_yolo.hdev has been adapted to changes in the ONNX reader, which can result in a change in the ordering of output dimensions for models loaded from ONNX files.
Note that this change affects the compatibility. Read more.add_dl_pruning_batch wrongly changed the batch normalization statistics for deep learning classification models containing batch normalization layers, which could lead to a worse model performance. This problem has been fixed.
There was an inconsistency between CPU and GPU results for depth-wise convolution layers on ARM architectures. The issue occurred when using a 5×5 kernel, stride 2, and non-zero padding. The problem has been fixed by switching to an alternative implementation for this specific configuration. Note that the alternative implementation can be about ten times slower for the affected cases.
The AI² optimization of a model containing a dropout layer with parameter ‘is_inference_output’ set to ‘true’ raised an error. This problem has been fixed.
The Deep Learning elementwise layer raised an error during training for the operation 'sum' with broadcasting. This problem has been fixed.
set_dl_model_param crashed for parameters expecting a handle if HNULL was passed. This problem has been fixed. Now, a corresponding error is raised.
The deep learning detection example dataset “screws” had wrong orientation labels for some screws of class_13. This problem has been fixed.
In some Deep Learning preprocessing procedures, the parameters 'image_range_min' and 'image_range_max' were restricted to -127 and 128. This applied in combination with the following model types: classification, multi-label classification, and detection. This restriction has now been removed. The procedure check_dl_preprocess_param in dl_preprocessing.hdpl has been updated accordingly.
The elementwise layer could produce incorrect results on runtime 'cpu' for the operations 'minimum', 'maximum', and 'sum' with broadcasting. This problem has been fixed.
Running CPU inference on a model containing a resized, broadcasting elementwise layer could cause a crash. This problem has been fixed.
In the context of Deep OCR training, the procedure gen_dl_samples did not consider the angle when calculating the cropped box dimensions, which could lead to cropped text in cases of strongly rotated input images. This problem has been fixed.
load_dl_model_weights did not properly update the weights of batch normalization layers with ReLU activation when the model runtime was set to 'gpu'. This problem has been fixed.
While training neural networks with the procedure train_dl_model, the evaluation of the model after a specific epoch could happen a single iteration too late. This problem has been fixed.
During the training of the neural networks, the learning rate was not displayed correctly when the learning rate changed. This problem has been fixed.
fit_dl_out_of_distribution in some cases leaked memory if it failed because the model could not correctly be applied to the samples in the dataset. This problem has been fixed.
Reading invalid legacy models with read_dl_model could lead to crashes. This problem has been fixed.
Switching the new alignment of the Deep OCR Recognition model across multiple threads on a GPU led to an error. This problem has been fixed.
set_text_model_param with 'separate_touching_chars' set to 'enhanced' crashed if the HALCON deep learning library could not be loaded. This problem has been fixed.
In the HDevelop example deep_ocr_workflow.hdev, the procedure display_recognition_alphabet did not show special characters if all capital letters were removed from the recognition alphabet. This problem has been fixed.
The error handling of set_deep_ocr_param and set_dl_model_param used together with the 'recognition_model' has been aligned. Both operators will throw the same errors if an invalid shape adjustment is recognized. Furthermore, it was possible to change the image height of the recognition model's input layer by bypassing the checks in set_deep_ocr_param via set_dl_model_param together with the parameter 'input_dimensions'. These problems have been fixed.
The cross entropy loss layer could crash when using invalid target class indices. This problem has been fixed.
When reading reduced domain JPEG 2000 images with four or more channels, HALCON always read the image as if it was saved with the full domain. This problem has been fixed.
When writing multi-channel images with more than three channels in JPEG 2000 format, HALCON saved the image as using the sRGB color space. Now, HALCON saves such images as having an unspecified color space.
Writing a lossless JPEG-XR image was not completely lossless. This problem has been fixed.
When reading PNG files that do not contain a .png file name extension in read_image, errors were not detected properly, potentially resulting in corrupted image data. This problem has been fixed.
Corrupted image files in PNM format have not always been detected. This problem has been fixed.
Using values exceeding the maximum image size for the read_sequence parameters SourceWidth and SourceHeight caused an error. This problem has been fixed.
On Windows, copy_file did not copy access rights and file timestamps of the source file. This problem has been fixed.
Under certain conditions, reading corrupted TIFF files caused a crash. This problem has been fixed.
In rare cases, when using large byte or uint2 images, many channels, and many threads, convol_channels could run out of memory, throw an out-of-memory exception, or crash. This problem has been fixed.
unwarp_image_vector_field did not handle multiple references to the same image correctly, for example, in the case of multi-object images, producing equal results for different vector fields. This problem has been fixed.
edges_sub_pix and edges_color_sub_pix returned interrupted edge contours in some cases, for example, in tight corners or when the edge points extracted by the edge operator were located in the same pixel. This problem has been fixed.
Note that this change affects the compatibility. Read more.highpass_image only processed the first channel for images of type 'byte'. This problem has been fixed.
sobel_amp('sum_sqrt') could crash for uint2 images. This problem has been fixed.
edges_sub_pix returned with an internal error when passing an image with specific, reduced domains in rare cases. This problem has been fixed.
convol_image led to an error for very large filter sizes and mirror margin. This problem has been fixed.
gauss_filter returned incorrect results for int4 images on 32-bit systems. This problem has been fixed.
gauss_filter handled inner and border region of 'real' images inconsistently. This problem has been fixed.
mean_image_shape in rare cases returned slightly incorrect results for byte images due to rounding problems on aarch64 with neon enabled. This problem has been fixed.
polar_trans_image_ext produced incorrect results for images of type 'real' if int_zooming was set to 'true' and the 'fused multiply-add' (FMA3) and 'Advanced Vector Extension 2' (AVX2) instruction sets were not enabled on x86 processors. This problem has been fixed.
draw_region handled sub-pixel coordinates incorrectly. Whenever the operator was executed on a strongly magnified image, the resulting region depended on which part of the pixel was actually touched: The lower-right quadrant of the image pixel resulted in the expected region pixel, whereas the other parts of the image pixel caused the resulting region to be moved by 1 to the left and/or up. This problem has been fixed.
In rare cases, get_data_code_2d_results could crash during print quality inspection for low quality Data Matrix codes when using 'quality_isoiec15415', 'quality_isoiec29158', 'quality_semi_t10_values', or their variants. This problem has been fixed.
When setting an AI² device with set_dl_model_param('device') for an advanced object detection model, the new error 7918 ("DL: The model supports this device only with optimize_dl_model_for_inference") is now returned instead of 7903 ("Error occurred while writing an ONNX model").
get_data_code_2d_objects unexpectedly returned an empty object for the module regions of a Data Matrix code in rare cases. This problem has been fixed.
find_data_code_2d in some cases leaked memory for the 2D code model QR Code if 'deformation_tolerance' was set to 'curved' and timeouts were enabled. This problem has been fixed.
find_data_code_2d could crash for 2D code models of symbol type “QR Code” if the persistence parameter was set to a value less than zero and the 'deformation_tolerance' parameter was set to 'curved'. This problem has been fixed.
In rare cases, get_data_code_2d_results returned wrong start/stop pattern grades of the print quality inspection of PDF 417 codes. This problem has been fixed.
For Data Matrix ECC 200 symbols, get_data_code_2d_results returned random values for 'structured_append' if queried for unsuccessful candidates. This problem has been fixed. Now, it always returns an empty tuple for 'structured_append' if queried for an unsuccessful candidate. This is true independent of the symbol type. I.e., the behavior is now consistent for Data Matrix ECC 200, QR Code, Micro QR Code, Aztec Code, and DotCode.
The Aztec Code reader in rare cases did not decode small symbols close to the image border. This problem has been fixed.
In rare cases, the Aztec Code reader returned the error 9205 ('Matrix is singular'). This problem has been fixed.
find_data_code_2d could show unexpected behavior if 'strict_quiet_zone' was set to 'yes' via set_data_code_2d_param and the symbol XLD of the candidate touched or crossed the image borders. In rare cases and under the mentioned circumstances, the strict quiet zone check may not have worked correctly. The issue affected all code types supporting 'strict_quiet_zone'. This problem has been fixed.
get_data_code_2d_objects returned wrong results for 'module_1_rois' and 'module_0_rois' for some Data Matrix codes with a square shape in rare cases. Additionally, get_data_code_2d_results returned wrong results for 'bin_module_data' for the same cases. These problems have been fixed.
The grid center determination used in the print quality inspection of the Data Matrix ECC 200 code reader had an issue caused by false edges close to the ends of the alternating pattern. This could affect both the robust and the reference decode algorithms. This problem has been fixed.
The print quality inspection for Aztec Codes sometimes returned a temporary memory error if the code to grade was very close to the image border. This problem has been fixed.
find_bar_code may have failed to decode certain readable barcodes in very rare cases. This problem has been fixed.
find_bar_code returned wrong results for CodeType = 'auto' and CodeType = ['Code 128', 'GS1-128'] in rare cases. This problem has been fixed.
find_bar_code could return wrong symbol regions and orientation values if 'min_identical_scanlines' was set to a value greater than one and multiple barcodes with an identical decoded string were covered by a successfully decoded scanline. In this case, the correspondence among the different scanline segments taken into account by the 'min_identical_scanlines' check could be wrong. This problem has been fixed.
find_bar_code in rare cases returned the error 3100 ("Wrong segmentation threshold"). This problem has been fixed.
find_bar_code in rare cases crashed when trying to decode GS1 DataBar Extended Stacked symbols. This may also have happened in the 'auto' mode of find_bar_code. This problem has been fixed.
find_bar_code crashed in rare cases during the quiet zone check. This only occurred with element_size_min below 2.0. This problem has been fixed.
find_bar_code sometimes returned errors regarding temporary memory. This problem has been fixed.
Some labels for print quality inspection according to ISO/IEC 15416 contained the term 'minimal' instead of 'minimum'. This problem has been fixed.
In rare cases, a successful decode of a barcode decoding attempt using the small_elements_robustness decode feature together with the quiet_zone check could lead to the unexpected error 3100 ("Wrong segmentation threshold"). This problem has been fixed.
find_bar_code may have failed to successfully decode certain readable barcodes in very rare cases. The issue affected only barcodes suffering from print growth or loss. This problem has been fixed.
find_bar_code crashed for CodeType 'GS1 DataBar Expanded Stacked' in very rare cases. This problem has been fixed.
find_bar_code could crash in very rare case. This problem has been fixed.
The training of the Barcode reader did not work with some codes that could be decoded with fixed parameters. This problem has been fixed.
For set_bar_code_param and set_bar_code_param_specific, the parameter 'quiet_zone' could not be set to 0. This problem has been fixed.
Some GS1 DataBar composites could not be decoded due to incorrect segmentation of the lower outline of the composite. This problem has been fixed.
HALCON could not read GS1 syntax dictionaries with "!" flags. Further, HALCON could not read GS1 syntax dictionaries with Windows CRLF line endings. These problems have been fixed.
gray_projections returned projections with inconsistent length and order in 'rectangle' mode when the smallest enclosing rectangle was axis-aligned. This problem has been fixed.
fit_surface_first_order returned wrong results if the input contained empty regions. This problem has been fixed.
find_ncc_model, find_ncc_models, find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, find_aniso_shape_models, find_generic_shape_model, and find_shape_model_3d failed to find some matches near the border of a reduced search image domain in rare cases. This problem has been fixed.
find_generic_shape_model, find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, and find_aniso_shape_models in special cases returned differing subpixel corrections for the matches depending on the architecture due to numerical differences. This problem has been fixed. In rare cases, a slowdown is to be expected.
find_generic_shape_model, find_shape_models, find_scaled_shape_models, and find_aniso_shape_models could return different matches when several models were used for a simultaneous search and at least one but not all of them were using max_deformation in combination with least-squares, compared with a single model search. This problem has been fixed.
find_generic_shape_model, find_shape_models, find_scaled_shape_models, and find_aniso_shape_models in specific cases calculated the overlap differently than in the single model find call, which resulted in differing matches in rare cases. This problem has been fixed.
find_shape_model, find_scaled_shape_model, find_aniso_shape_model, and find_generic_shape_model in very rare cases suppressed matches when many threads were used. This problem has been fixed.
find_generic_shape_model, find_shape_models, find_scaled_shape_models, and find_aniso_shape_models may have returned slightly differing matches than the ones found by a single model find call due to numerical issues. This problem has been fixed.
create_ncc_model, find_ncc_model, find_ncc_models, and determine_deformable_model_params could take a long time to set the parameter values and could not be interrupted during this time. This problem has been fixed.
set_generic_shape_model_param allowed setting values larger than 32 for max_deformation, contrary to other shape-based matching operators. This problem has been fixed.
create_ncc_model using the NEON implementation could access wrong memory. This problem has been fixed.
find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, find_aniso_shape_models, and find_generic_shape_models did not consider max_deformation values in the score calculation when called without least-squares refinement. This problem has been fixed.
create_shape_model, create_scaled_shape_model, create_aniso_shape_model, and train_generic_shape_model added the last model point twice (or even three times) to the model for some rare combinations of the number of model points and optimization. This problem has been fixed. Note that as a consequence, the score of matches with new models may change and hence different matches may be found.
find_generic_shape_model and find_aniso_shape_model had slowdowns during tracking. This problem has been fixed.
find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_model, find_aniso_models, find_generic_shape_model, and determine_shape_model_params could take a long time to set the parameter values and could not be interrupted during this time. This problem has been fixed.
find_shape_models, find_scaled_shape_models, find_aniso_models, and find_generic_shape_model returned incorrectly sorted matches, in case different values were used for the lowest pyramid level for the different passed models. This problem has been fixed.
create_ncc_model added a rotation step more than necessary and hence increased the angle width. This problem has been fixed.
find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, find_aniso_shape_models, and find_generic_shape_model could return a memory error if 'model_cache' was set to 'false', 'subpixel' to 'least_squares', and the match instance set such that the model was very close to the border. This problem has been fixed.
train_generic_shape_model overwrote the parameter 'model_cache', setting it always to 'true'. This problem has been fixed.
find_generic_shape_model, find_shape_model, find_shape_models, find_scaled_shape_model, find_scaled_shape_models, find_aniso_shape_model, and find_aniso_shape_models did not consider max_overlap correctly due to an incorrect computation of the smallest enclosing rectangle in very rare cases. This problem has been fixed.
set_metrology_object_fuzzy_param could crash when given invalid parameter values. This problem has been fixed.
measure_projection could return values differing more than one gray value when compared to its values with another system setting for int_zooming. This problem has been fixed.
When performing measurements on int4 images with very large pixel values, overflows could occur. This problem has been fixed.
The metrology model parameter 'scale' and the metrology object parameter 'distance_threshold' were previously not always serialized depending on which other model parameters were set. This problem has been fixed.
hinfoxl attempted to load the standard HALCON library, not the XL version. This problem has been fixed.
For x64-linux, the HALCON library exported symbols from the Intel compiler runtime libraries. This problem has been fixed.
The HGetCElementHN macro was defined incorrectly. This problem has been fixed.
get_grayval_interpolated could crash or return incorrect results when used on very large images in HALCON XL. This problem has been fixed.
Code analysis detected uninitialized values. This problem has been fixed.
Due to the refactoring of the code, extension packages using the functions HRLClip2, HRLRelClip, or HRLClip are no longer source-code compatible and must be adapted.
Even when the region had enough capacity, HALCON could incorrectly signal H_ERR_NOT_ENOUGH_CHORDS, leading to incorrect region generation in affected cases. This problem has been fixed. The reliability of operators such as gen_region_polygon_filled and gray_inside has been improved.
Note that this change affects the compatibility. Read more.channels_to_image and other operators performing region intersections internally could have failed for very small regions of 1x1 or 2x2 pixels. This problem has been fixed.
region_features(Regions : : Features : Value) crashed for multiple features when computed on a region that extends to the limits of the available coordinate range. This problem has been fixed.
segment_image_mser could incorrectly suppress a valid region when 'may_touch_border' was set to 'false'. The problem occurred for regions that do not touch the border, but were an interior region of a larger, enclosing region that touches the border. This problem has been fixed.
watersheds_threshold and watersheds_threshold crashed in very rare cases. This problem has been fixed.
intersection_contours_xld did not return any intersection point if one of the passed contours was a single, unique point lying on the other passed contour. This problem has been fixed.
intersection_contours_xld used to throw a dispensable low-level error message. This error message has been removed.
Under certain circumstances hinfo could be manipulated into loading two different HALCON shared libraries simultaneously, leading to inconsistent output. This problem has been fixed.
For affine_trans_image, the parameter 'AdaptImageSize' wrongly accepted string values other than 'true' and 'false'. In this case, the value 'false' was assumed. This problem has been fixed. Now, the operator returns an error if another value than 'true' or 'false' is used.
HALCON implements copy_dict, write_dict, and serialize_handle for dict handles recursively. This can lead to stack overflow if the dict is nested too deeply, so HALCON will now report an error if any of these operators is called with a dict that is nested to a depth of more than 100.
tuple_number treated floating point numbers starting with a plus sign as strings, not numbers. This problem has been fixed.
tuple_sum and tuple_cumul could return incorrect results for very large integer input values. This problem has been fixed.
tuple_cumul did not return all values as floating-point values if the input contained at least one floating-point value, as described in the reference documentation. This problem has been fixed.
union_adjacent_contours_xld produced in incorrect unions of adjacent contours when closed contours were involved. This problem has been fixed.
When using intersection_region_contour_xld in HALCON XL with large coordinates, the returned intersections were not always correct. This problem has been fixed.
For HALCON/C++ and HALCON/C, the HBase.h header file redefined the TRUE and FALSE macros to true and false, respectively. This could lead to compiler warnings on Windows. Now, HBase.h will no longer redefine TRUE or FALSE if the HC_NO_LEGACY_TYPES macro is defined.
In some cases, a HALCON/C application could crash when destroying an application thread previously used to execute HALCON operators. A necessary condition was that at least one operator, typically the last one executed in the thread, returned an error result. This problem has been fixed.
The signature of the receive_data operator was wrong in the HALCON language interfaces. As a result, receive_data could not be used from the Python language interface. This problem has been fixed.
The Python qt_gui example did not build because of an outdated reference to version 5.15.2 of PyQt5. This problem has been fixed.
A warning was displayed when building the C++/CLI example "interoperate". This problem has been fixed.
HALCON Variable Inspect Extension for Visual Studio is now also available for Microsoft Visual Studio 2026. It can be installed with the installer “Variable Inspect Extension for Visual Studio 2022 and later”.
Debugging C++ tuples in older HALCON versions with a new HALCON Variable Inspect version was not possible anymore. This problem has been fixed.
The HALCON Variable Inspect Extension was inconsistent about when it was responsive or not. Now, it is only responsive during debugging and while at a break, as intended. Further, on first open, the variables to inspect were not loaded correctly. These problems have been fixed.
HALCON Variable Inspect is now clearly visible as being disabled when not in debug mode.
User settings were not stored correctly. This problem has been fixed.
The Easy Extensions Interface now uses the HALCONEXTENSIONS environment variable to determine the location of configuration file and plugins at runtime. This means the built adapter can be transferred to a different machine more easily. Previously, the absolute build path was compiled into adapter_dotnet.dll.
In addition, the ManagedLibrary.runtimeconfig.json file has been moved into the dotnet folder (next to ManagedLibrary.dll).
The CMake build for the Easy Extensions Interface, as well as the build instructions in the corresponding Readme file have been improved to simplify the build experience.
hcomp could insert invalid characters into the HALCON help files under certain circumstances. This problem has been fixed.
hbench did not run the complete set of benchmarks but produced an error when the -operator option was used and no individual operator was specified. This problem has been fixed.
The latest information about new interface revisions and newly supported image acquisition devices can be found on MVTec's web server. Please refer to the release notes within the documentation of the individual image acquisition interfaces for information about improvements, bugfixes, or whether a new revision of the corresponding device driver is required.
Image Sources now support persisting either individual parameter groups or the complete Image Source configuration, enabling easy restoration of device and setup parameters across sessions and machines.
Image Sources now support querying a nested dictionary of each parameter group containing all parameter categories, sub-categories, and their parameters.
For HALCONARCH x64-linux, libmvimgsrc.so contained undefined references to Intel runtime libraries. This problem has been fixed.
Some GenTL Producers refused to open a module without a previous call to TLUpdateInterfaceList or IFUpdateDeviceList. If this case is detected, these functions are now called, and opening the module is retried.
A new Image Source plugin, mvtec_file, has been added. It allows simulating a real application with image files when a real camera is not desired or available. The first version only supports the JPEG and PNG formats.
The range of compatible versions of the GenICam GenTL standard has been extended. Now, all versions 1.3 and newer are supported.
The new image acquisition now supports GenICam Chunk data. It can be obtained with get_image_source_param, the same way known from get_framegrabber_param. The Chunk data is additionally contained in the Data dictionary returned by fetch_from_image_source and snap_from_image_source. The usage of Chunk data is demonstrated in the new example hdevelop/ImageSource/image_source_chunks.hdev.
Reading unavailable registers through get_image_source_param has been improved to deal with possible unexpected problems when accessing the register length.
The latest information about new interface revisions and newly supported digital I/O interfaces can be found on MVTec's web server. Please refer to the release notes within the documentation of the individual digital I/O interfaces for information about improvements, bugfixes, or whether a new revision of the corresponding device driver is required.
The term “data code” has been replaced with the more commonly used term “2D code” throughout the documentation.
The individual Image Source Plugins are now documented online. The pages are linked from the operator reference of the Image Source operators.
The information about supported .NET versions was outdated and has been clarified.
The documentation did not mention that Variable Inspect does not support remote debugging. This problem has been fixed.
The documentation of create_caltab has been expanded to include information about the example calibration plate description files provided with HALCON, as well as the calibration plate portfolio available on the MVTec website.
The reference manual entry of find_data_code_2d did not clearly state that the parameter 'stop_after_result' needs to be set to a value greater than 0. This problem has been fixed.
The documentation of register_object_model_3d_pair did not clearly describe the interaction between default_parameters and additional parameter settings. This problem has been fixed.
The reference manual entries of all operators whose results are affected by the set_system parameter 'clip_region' were updated with a note indicating the clipping behavior.
The reference manual entry of the operators find_aniso_shape_models, find_scaled_shape_models, find_shape_models and find_generic_shape_model did not mention the special behavior of the parameter MinScore or 'min_score', respectively, on the highest pyramid level in a multi-model case. The documentation has been adapted accordingly.
The reference manual entry of get_data_code_2d_param falsely stated that only for DotCode symbols the module size restrictions are not checked even if 'strict_model' is set to 'yes'. However, this is the case for all symbol types.
The reference manual entry of set_data_code_2d_param falsely stated that the parameter 'abort' is ignored when running in training mode.
These problems have been fixed.
The default values for 'detection_optimize_for_inference' and 'recognition_optimize_for_inference' have been removed from the documentation of get_deep_ocr_param, since the actual value depends on the used underlying model.
The reference manual entry of get_generic_shape_model_param did not state that multi-value GenParamValues cannot be returned together with further values and therefore need to be queried in a distinct get_generic_shape_model_param call. The documentation has been adapted accordingly.
The reference manual entry of set_dl_model_param stated that the parameter 'alignment_available' could be set for OCR recognition models, although this was not possible. This problem has been fixed, and the corresponding documentation entry has been removed.
The Deep 3D Matching documentation and examples have been extended with some helpful information on the usage of the method. The reference manual entry of create_deep_matching_3d now provides a simple example on how to set symmetries of the 3D object model. The reference manual entry of get_deep_matching_3d_param now mentions the importance of having camera_pose in the same unit as the 3D object model.
The reference manual entry of get_dl_model_param did not mention the parameter 'inference_outputs'. This problem has been fixed.
The reference manual entry for set_spy mentioned the option “error” although the option had no effect. This problem has been fixed. The outdated entry has been removed.
The reference manual entry of the signature for the callback used with the 'do_low_error' set_system parameter was incorrect. This problem has been fixed.
For some operators, the output of get_operator_info contained invalid UTF-8 characters and incorrect output of quotation marks. This problem has been fixed. The following operators were affected by invalid UTF-8 output:
triangulate_object_model_3d, reconstruct_surface_stereo, set_stereo_model_param, get_dl_model_param, get_deep_ocr_param
The system requirements now state how using Windows IoT systems can affect HALCON functionality.
The Installation Guide now warns about changing the system time when using temporary licenses.
The documentation incorrectly stated that dict access expressions on long tuples can be slow due to tuples being copied. The documentation has been updated, accessing long tuples is no longer slow.
The files HalconUninst.dll and hcheck_cpu.exe in misc/x86-win32 were not used anymore in newer versions of HALCON and have been removed from the file set, including the containing folder.
The installation of the package "Deep Learning Core" erroneously did not depend on the installation of "Microsoft Visual C++ Redistributable". This problem has been fixed.
The Microsoft Visual C++ redistributable installed with HALCON has been updated to version v14.44.35211.0. This version supports all applications created using Visual Studio 2015, 2017, 2019, or 2022.
When run on a system under very high load, HALCON could erroneously detect system-time manipulation. This problem has been fixed.
For aarch64-linux, the CID licensing has been extended to also accept NVME SSD serial numbers.
hinfo now reports the full version number of HALCON, including the hotfix version.
The third-party library onnx has been updated to version 1.22.0.
The third-party library libjpeg-turbo has been updated to version 3.2.0.
The third-party library libcurl has been updated to version 8.22.0.
The third-party library mimalloc has been updated to version 2.5.2.
HALCON now uses version 8.41a of the CodeMeter Runtime.
The third-party library libcurl has been updated to version 8.19.0.
The third-party library mimalloc has been updated to version 2.2.7.
The Cycles library used in the Scene Engine has been updated to version 4.4.0.
mimalloc has been updated to version 2.2.4.
HALCON now uses version 8.30a of the CodeMeter Runtime.
LibTIFF has been updated to version 4.7.1.
Follow this link to read about the changes of previous HALCON versions.