Release Notes for HALCON 24.11.1.0 Steady

Classification

• classify_class_knn is now internally parallelized over the input points, using the automatic operator parallelization.

• A new deep learning method named multi-label classification has been added to HALCON. This method allows assigning multiple classes to an image at once, with little effort to label the data and a lower inference time compared to other model types like detection.
  In addition to extending the deep learning operators and procedures with the new multi-label classification model type, the following procedures are introduced:
  create_dl_model_multi_label_classification, read_dl_dataset_multi_label_classification, convert_classification_dataset_to_multi_label_classification, convert_detection_dataset_to_multi_label_classification, convert_segmentation_dataset_to_multi_label_classification, dev_display_multi_label_classification_evaluation

  The HDevelop example program hdevelop/Deep-Learning/MultiLabelClassification/dl_multi_label_classification_workflow.hdev has been added to show how the DL model of multi-label classification can be used in HALCON.

  A chapter describing the use of the new multi-label classification method has been added to the reference manual ("Deep Learning" > "Multi-Label Classification").
  The new procedure dev_display_multi_label_classification_evaluation can be used to visualize the results of a multi-label classification model.
  An existing classification, recognition, or segmentation dataset can be converted into a multi-label classification dataset using the provided conversion procedures.

• The NVIDIA CUDA Basic Linear Algebra Subroutine library (cuBLAS) and NVIDIA CUDA Deep Neural Network library (cuDNN) have been updated to the latest versions. To support also the latest NVIDIA graphics cards, HALCON now ships and supports these libraries for CUDA versions 12.1 on x64-linux and x64-win64 platforms. The libraries can be found in the cuda12_1 subdirectories of the thirdparty library directory.
  The following versions of the NVIDIA libraries are shipped for CUDA 12.1:

  • NVIDIA CUDA Basic Linear Algebra Subroutine library (cuBLAS) from the CUDA 12.1 Toolkit version 12.1.3.1
  • NVIDIA CUDA Deep Neural Network library (cuDNN) version 8.9.2

  For aarch64-linux, HALCON now supports NVIDIA Jetpack 5.1.1.

  Note that this change affects the compatibility. Read more.

• The DL Framework now contains the new operator create_dl_layer_matmul. This allows building custom CNN graphs that contain matrix multiplications of two inputs DLLayerA and DLLayerB, where the 2D matrices are encoded in the last two dimensions (H, W) of the input tensors, respectively.

• create_dl_layer_elementwise now allows creating a DL layer with a new elementwise operation type 'division'.

• The DL framework now contains the new operator create_dl_layer_reduce that creates a Reduce layer within a custom DL model. The layer can be used to reduce an input tensor along one or several given axes by a given operation. Currently, the operation 'norm_l2' to compute the L2-Norm along the given axis is supported.

• The anomaly detection performance of Global Context Anomaly Detection (GC-AD) has been improved. This is enabled by a novel loss layer in the local network of the GC-AD model and by new pretrained model weights. The updated model has the same name as in previous HALCON versions: pretrained_dl_anomaly_global_context.hdl Although it may perform worse in some cases, this model detects anomalies considerably better on average than the previous versions of GC-AD.
  The changes to the GC-AD model do not affect the runtime and memory requirements during inference. The novel loss layer has a small effect on the training duration. For the default image size and patch size setting, this effect is only marginal. Another effect of the updated loss layer is that the loss values observed during training are expected to be numerically higher.

• optimize_dl_model_for_inference has been extended to support int8 precision together with the Deep Counting model.

• A new operator create_dl_layer_reduce with operations sum and norm_l2 has been introduced.

• The distance loss layer now supports the L1 distance. create_dl_layer_loss_distance has been extended accordingly.

• The Deep Learning Inference module has been renamed Deep Learning Enhanced and the Deep Learning Training module has been renamed Deep Learning Professional. For more information on which technology belongs to which module, see the Installation Guide (chapter 5.4.2).

• The required license modules have been adapted according to the new deep learning licensing.

• get_dl_model_param has been extended with the parameter 'min_version'. It returns the minimal required HALCON version supporting the given model. This allows users to ensure their models are compatible with older HALCON versions.

• The runtime initialization of a GPU device on a deep-learning model took a large amount of time since HALCON 23.11. This has been significantly improved.

  Note that this change affects the compatibility. Read more.

• HALCON has been expanded to include the Out-of-Distribution Detection (OOD) function for all deep learning classification models. This function recognizes when an input is outside the distribution of the training data.

  In addition to the extension of the deep learning operators and procedures with the new out-of-distribution detection function, the operator fit_dl_out_of_distribution and the procedure add_dl_out_of_distribution_data have been introduced.

  The HDevelop example program hdevelop/Deep-Learning/Classification/detect_out_of_distribution_samples_for_classification.hdev has been added to show how to set up a classification DL model with this feature.

  The reference manual entry "Deep Learning” > “Classification” has been extended with instructions on how to fit a classification model for out-of-distribution detection.

  The procedure evaluate_dl_model has been extended to evaluate the out-of-distribution detection feature using multiple evaluation metrics.

• read_dl_model now supports reading ONNX models that contain Transpose operations.

• read_dl_model now supports reading ONNX models that contain Reshape operations.

• HALCON has been extended with Deep Counting. This method makes it possible to count objects in an image easily and robustly without setting complex parameters.
  The operators for Deep Counting are:

  • apply_deep_counting_model
  • create_deep_counting_model
  • prepare_deep_counting_model
  • set_deep_counting_model_param
  • get_deep_counting_model_param
  • write_deep_counting_model
  • read_deep_counting_model

  The HDevelop example programs hdevelop/Matching/Deep-Counting/deep_counting_workflow.hdev and hdevelop/Matching/Deep-Counting/deep_counting.hdev have been added to show how to use Deep Counting in HALCON.
  The example hdevelop/Matching/Deep-Counting/deep_counting.hdev has been extended with the new images flasks_[01-03].jpg, pipes_[01-03].jpg, and rods_[01-03].jpg from the subdirectory images/counting.
  The example hdevelop/Matching/Deep-Counting/deep_counting_workflow.hdev uses the already existing images, bottle_crate_[01-24].png, from the subdirectory images/bottles.

  The new procedure draw_deep_counting_templates can be used to draw templates for the Deep Counting model. The new procedure dev_display_deep_counting_results can be used to visualize the results of Deep Counting.

  The reference manual has been extended with a chapter describing how to use the new Deep Counting functionality ("Matching" > "Deep Counting").

• read_dl_model can now read ONNX models that contain a ConvTranspose operation. Moreover, the transposed convolution (create_dl_layer_transposed_convolution) does now support a bias term and CPU training.

• HALCON has been extended with Deep OCR detection training. Now, the detection component of a Deep OCR model can be retrained on a specific application domain for even better detection results.

  The new HDevelop example hdevelop/OCR/Deep-OCR/deep_ocr_detection_training_workflow.hdev has been added to show how to use this feature in HALCON. For this example, a new set of images has been added. It is located in the example images folder named "punched_numbers".
  The existing HDevelop example program hdevelop/OCR/Deep-OCR/deep_ocr_prelabel_dataset.hdev has been adapted and it shows how to prelabel images using Deep OCR and create a dataset that can be imported into the Deep Learning Tool for refining the labels.

  The following deep learning procedures have been extended:

  • augment_dl_samples
  • create_dl_preprocess_param
  • create_dl_preprocess_param_from_model
  • dev_display_dl_data
  • dev_display_dl_data_tiled
  • evaluate_dl_model
  • preprocess_dl_samples
  • preprocess_dl_model_images
  • preprocess_dl_model_augmentation_data
  • train_dl_model

  The following new procedures are now available:

  • convert_ocr_detection_result_to_object_detection
  • dev_display_dl_invalid_samples
  • preprocess_dl_model_images_ocr_detection
  • read_dl_dataset_ocr_detection

  The operators set_dl_model_param and get_dl_model_param have been extended with Deep OCR detection specific parameters:

  • 'min_character_score'
  • 'min_link_score'
  • 'min_word_area'
  • 'min_word_score'
  • 'orientation'
  • 'sort_by_line'
  • 'tiling'
  • 'tiling_overlap'

  The reference manual chapters "Deep Learning", "Deep Learning/Model", and "OCR/Deep-OCR" as well as the operator entry for get_dl_model_param, set_dl_model_param, set_deep_ocr_param, and get_deep_ocr_param have been extended. They now describe how to use the method and the corresponding parameters. The chapter on Deep OCR in the Solution Guide I Basics has been adapted accordingly.

• The pretrained Deep OCR detection model has been improved and extended. The resulting detection boxes of apply_deep_ocr for the modes ‘detection’ and ‘auto’ are now more precise.
  Additionally, a new compact detection component has been added. This compact model is slightly less precise than the default. As a benefit, it uses significantly less memory and is much faster during inference. Both models can be retrained using Deep OCR detection training.
  set_deep_ocr_param has been extended with the option to specify a filename, ‘compact’ or ‘default’ for the parameter ‘detection_model’. The parameter ‘recognition_model’ has been extended with the option to set a filename or ‘default’.
  The following pretrained model has been updated:
  dl/pretrained_deep_ocr_detection.hdl
  The following pretrained model has been added:
  dl/pretrained_deep_ocr_detection_compact.hdl

  Note that this change affects the compatibility. Read more.

• It is now possible to set certain model parameters after calling optimize_dl_model_for_inference. This eliminates the need to repeat the time-consuming optimization.
  An optimized model's architecture cannot be changed, therefore only a selection of parameters is supported. In addition, restrictions may be placed on the supported parameters to meet these constraints.
  The following parameters of each model type can be changed using set_dl_model_param after model optimization:

  • Anomaly Detection: standard_deviation_factor
  • Global Context Anomaly Detection: anomaly_score_tolerance
  • Classification: class_names
  • Detection: class_names, max_num_detections, max_overlap, max_overlap_class_agnostic, min_confidence
  • Deep OCR Detection Component: min_character_score, min_link_score, min_word_score, orientation, sort_by_line, tiling, tiling_overlap
  • Deep OCR Recognition Component: alphabet, alphabet_internal, alphabet_mapping
  • Segmentation: class_names

  set_deep_ocr_param has also been extended with this new functionality and supports the following parameters after optimization:

  • Deep OCR: detection_min_character_score, detection_min_link_score, detection_min_word_score, detection_orientation, detection_sort_by_line, detection_tiling, detection_tiling_overlap, recognition_alphabet, recognition_alphabet_internal, recognition_alphabet_mapping

• set_dl_model_param and get_dl_model_param have been extended. Using the parameter 'image_size' the width and height of the input image can be set and get, while keeping the depth, the same.

File

• Memory block support has been improved. Previously, handling of empty memory blocks was not supported. Now, memory blocks can be handled even when empty.

  Note that this change affects the compatibility. Read more.

• list_files has been extended to support listing of symbolic links on Linux systems.

• HALCON has been extended with serialization and deserialization support for the data types 'serialized_item', 'memory_block' and 'encrypted_item'.

Graphics

• set_lut (and its HDevelop counterpart dev_set_lut) have been extended, such that the minimum and maximum gray value for the mapping of the LUT can be specified by passing a tuple [LUTName, MinGray, MaxGray] to this operator.

• Skia, the library used for HALCON's visualization, has been updated to milestone 115.

• disp_text and dev_disp_text have been extended with new generic parameters:

  • `backdrop_blur_sigma' to specify the amount of blur of the background of the text box
  • 'rotate_phi', 'rotate_col', and 'rotate_row' to enable rendering of rotated text
  • 'border_width' and 'border_color' to specify how the border of the text box is rendered

• Legacy gnuplot operators have been removed from HALCON.

  Note that this change affects the compatibility. Read more.

• The new LUTs 'batlow' and 'invers_batlow' have been added. These LUTs are perceptually uniform and color-vision-deficiency friendly.

• set_drawing_object_params has been extended with the parameter 'marker_size' to make the size of the anchors used to manipulate drawing objects configurable.

Image

• HALCON has been extended with the operator add_image_border. It can be used to add an arbitrarily sized border with a constant gray value to an image.

Matching

• train_generic_shape_model now offers the possibility to determine the parameter 'min_score' by extended parameter estimation. Doing so, a suitable value is estimated given representative matching samples. Accordingly, set_generic_shape_model_param has been extended to set representative samples to the matching model.

• Component-based matching is set to legacy for HALCON 24.05.

• The operators for finding shape models for shape-based matching 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 are now faster on processors that support the SSE2, AVX2, AVX512, or NEON instruction sets. The speedup depends on the configuration of the shape-based matching (e.g., on the number of pyramid levels, the sub-pixel refinement, the available instruction set, and highly on the value for greediness) and can therefore be non-existent as well. The following speedups can be expected for an exemplary configuration:

  Image Type	Available Instruction Set	Speedup
  byte, uint2	SSE2	15%
  	AVX2	30%
  	AVX512	50%
  	NEON	20%

  As a consequence of this speedup, the speed of find_component_model and find_shape_model_3d has also improved for some configurations.

• train_generic_shape_model now offers a new method to determine search parameter values by extended parameter estimation. Given representative matching samples, the estimation result replaces 'greediness' and 'min_score'. Accordingly, set_generic_shape_model_param has been extended with a new parameter named 'extended_parameter_estimation'. The new method can be enabled with the value 'per_level'.
  As a consequence, the extended parameter estimation of 'min_score' is now enabled by setting the new parameter to the value 'linear'. The HDevelop example train_generic_shape_model_sample_workflow.hdev has been adapted.
  As the estimation replaces 'greediness' and 'min_score', the result values cannot be retrieved by get_generic_shape_model_param.

• For train_generic_shape_model, determining 'min_score' through the extended parameter estimation method 'linear_estimate' has been modified. The criterion for accepting matches in samples has been improved. This may change the behavior in rare cases.
  The operator now also returns the error 8684 ("Extended parameter estimation failed") in case of no or too many found matches on a sample.

• The generic shape matching interface now allows evaluation of the max overlap for an instance for all models passed in find_generic_shape_model. To enable global max overlap, call set_generic_shape_model_param with 'max_overlap_global_enable' set to the value 'true'.

Miscellaneous

• The 'val' parameter of the HPGetPPar function was changed from Hcpar* H_RESTRICT to Hcpar const* H_RESTRICT*.

  Note that this change affects the compatibility. Read more.

• The Hpar f union was declared legacy since HALCON 11.0 and has been removed. Furthermore, the macro FLOAT_PAR has been removed.

  Note that this change affects the compatibility. Read more.

• The minimum system requirements for HALCON library on armv7a-linux, x64-linux platform, and armv8-linux platforms have been increased to gcc 7 and glibc 2.27.

  Note that this change affects the compatibility. Read more.

OCR

• HALCON has been extended with an operator tuple_str_distance that calculates the edit distance between strings.

Parallelization

• HALCON's automatic operator parallelization can now use more than 64 cores when running Windows 11 or Windows Server 2022.

Region

• The Extension Package Programmer's Manual now explains more details of the HSetExtendedErrorInfo function and also mentions the related HSetExtendedErrorInfoF function.

• On August 24, 2022, a security vulnerability found in 6th through 11th generations of consumer and 1st through 4th generations of Xeon Intel x86-64 microprocessors has been reported to INTEL (CVE-2022-40982). To resolve the vulnerability, microcode updates have been released that significantly reduce the performance of HALCON operators that make use of gather instructions. To mitigate the performance impact of the microcode updates, many affected HALCON operators have been revised.

  On startup, HALCON now automatically detects if the used CPU is affected and prefers the usage of alternative implementations without gather instructions. To query the recommendation of HALCON, get_system has been extended with the new parameter 'avx2_gather_recommended'. If the old implementations with gather instructions is desired (e.g., for unpatched CPUs), set_system now has the new parameter 'avx2_gather_enable' to control whether AVX2 gather instructions should be used or not. 'avx2_gather_enable' is also available for get_system to query the current status.
  HALCON automatically sets 'avx2_gather_enable' to 'false' for AMD CPUs Zen1 and Zen2 because of their poor performance with gather instructions.

  The following operators are affected by the security vulnerability and have been revised to either mitigate the performance impact or improve the performance compared to previous HALCON versions:

  • map_image is now faster for maps of type 'bilinear‘ and 'coord_map_sub_pix‘ for byte and uint2 images. The slowdown on patched CPUs (up to −70%) has been compensated completely and a speedup of up to 25% (byte) and 30% (uint2) compared to previous HALCON versions can be expected.
    For unpatched CPUs that enable the gather instructions or for unaffected CPUs, a speedup of up to 40% can be expected. This only applies to image sizes/maps that do not require HALCON XL.
  • affine_trans_image, affine_trans_image_size, zoom_image_factor (if the 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' and system parameter 'int_zooming' set to 'true' have been revised for interpolation 'bicubic' for all available image types ('byte', 'int2', 'uint2', and 'real') on processors that support AVX2 to mitigate the −70% slowdown.
    For images of type 'byte' and 'int2', the operators are up to 20% faster, for images of type 'uint2', the operators are up to 25% faster and for images of type 'real', a speedup of up to 40% can be expected when compared to previous HALCON versions for patched CPUs. For unaffected or unpatched CPUs, a speedup of up to 5% for images of type 'int2', a speedup of up to 10% and for images of type 'real', a speedup of up to 15% can be expected when compared to previous HALCON versions.
  • affine_trans_image, affine_trans_image_size, zoom_image_factor (if the 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 'true' and system parameter 'int_zooming' set to 'true' have also been revised for interpolations 'bilinear', 'constant', and 'weighted' on processors that support AVX2 to mitigate the up to −70% slowdown. The same holds for gen_gauss_pyramid using interpolation 'constant' and 'weighted' for images of type 'uint2', when the Scale is not exactly 0.5. The slowdown of images of type int2 and uint2 could not be compensated entirely and are now up to −20% slower than with an unpatched CPU. For images of type real, the performance is mostly equal for patched CPUs, with a slowdown of up to −15%.
  • zoom_image_factor (if the ScaleWidth and ScaleHeight are exactly 0.5), zoom_image_size (if the output image size is half the input image size) have been revised for interpolation method 'bicubic', 'bilinear', and 'none' on processors that support AVX2 to mitigate the up to −65% slowdown. HALCON 24.05 is now faster by up to 20% for images of type 'real' compared to previous versions. For newer Intel generations that are not affected, the speedup can be up to 30%.
  • rotate_image (for angles of 90 and 270 degrees) has also been revised for all supported interpolation types for images of type 'byte' on processors that support AVX2 to mitigate the up to −65% slowdown. Even with an installed microcode update, the performance is identical or slightly faster when compared with an unpatched CPU.
  • projective_trans_image and projective_trans_image_size with system parameter 'int_zooming' set to 'true' have also been revised for interpolations 'bilinear' on processors that support AVX2 to mitigate the up to −55% slowdown.
    The slowdown of images of type uint2 could not be compensated entirely, and is now up to −20% slower than with an unpatched CPU. For images of type real, the slowdown is slightly less and is up to −10% slower than without a microcode update.
  • polar_trans_image_ext with system parameter 'int_zooming' set to 'true' has been revised for interpolation 'bilinear' for images of type 'int2' and 'uint2' on processors that support AVX2 to mitigate the −50% slowdown. Although the slowdown could not be compensated entirely, the performance regression is now only up to –20%.
  • derivate_gauss has been revised for all supported image types and the performance regression of up to −20% due to a microcode update could almost entirely be fixed with a remaining slowdown of about −5% for images with full domain. As a result of this, edges_color, edges_color_sub_pix and edges_sub_pix for images of type 'byte' and 'uint2' for Filter 'canny', points_foerstner for Smoothing 'gauss', critical_points_sub_pix for Filter 'gauss' and 'facet' and all supported image types, local_max_sub_pix and local_min_sub_pix for Filter 'gauss' and 'facet' and all supported image types and saddle_points_min_sub_pix for Filter 'gauss' and 'facet' have been affected in an identical or slightly less manor.
  • find_planar_calib_deformable_model has been revised on processors that support AVX2 to mitigate the up to −10% slowdown. The slowdown could be compensated entirely.
  • gen_grid_rectification_map has been revised for images of type 'byte' and 'uint2' on processors that support AVX2 to mitigate the up to −45% slowdown. Although the slowdown could not be compensated entirely, the performance regression is now only up to –20%.
  • binary_threshold has been revised for images of type 'byte' on processors that support AVX2 for Method 'smooth_histo' to mitigate the up to −45% slowdown. Although the slowdown could not be compensated entirely, the performance regression is now only up to –20%.
  • histo_to_thresh has been revised for absolute histograms with 256 values on processors that support AVX2 to mitigate the up to −60% slowdown. The slowdown could not be compensated entirely, the performance regression is now only up to –30%.
  • determine_shape_model_params and determine_ncc_model_params have been revised for images of type 'uint2' on processors that support AVX2 to mitigate the up to −10% slowdown. The slowdown could be compensated entirely.
  • create_shape_model, create_scaled_shape_model and create_aniso_shape_model have been revised for Templates of type 'uint2' on processors that support AVX2 to mitigate the up to −20% slowdown. HALCON 24.05 is now faster by up to 15% compared to previous versions.

  Although not affected by the security vulnerability, many HALCON operators have been sped-up:

  • affine_trans_image, affine_trans_image_size, zoom_image_factor (if the 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 'true' with system parameter 'int_zooming' set to 'true' for interpolation 'bilinear', 'constant', and 'weighted', on processors that support AVX2 are now up to 35% faster for affected CPUs and up to 60% faster for newer, unaffected CPUs for images of type 'byte'. The same holds for gen_gauss_pyramid using interpolation 'constant' and 'weighted' for images of type 'byte', 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 affected CPUs and up to 65% faster for newer, unaffected CPUs for interpolation 'constant' on processors that support AVX2.
  • 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 AVX2 are now up to 70% faster for affected CPUs and up to 60% faster for newer, unaffected CPUs for images of type 'byte'.
  • gen_bundle_adjust_mosaic and gen_projective_mosaic are now up to 70% faster for images of type 'byte' for affected and patched CPUs while the speedup for newer, unaffected CPUs is up to 60%, when compared to HALCON 23.05.
  • A speedup of 10% on average can be expected for the robust PQI, i.e., for the print quality inspection if 'quality_isoiec15415_decode_algorithm' is set to 'robust' using set_data_code_2d_param.
  • polar_trans_image_ext with system parameter 'int_zooming' set to 'true' for interpolation 'bilinear' on processors that support AVX2 are now up to 35% faster for affected CPUs and up to 45% faster for newer, unaffected CPUs for images of type 'byte'.
  • create_shape_model, create_scaled_shape_model, and create_aniso_shape_model are now up to 20% faster for affected CPUs and up to 25% faster for newer, unaffected CPUs on processors that support AVX2 for Templates of type 'byte'.

• The following operators now support parameter broadcasting, where input tuples of length 1 are automatically extended and used for all created elements:
  gen_circle, gen_circle_sector, gen_ellipse, gen_ellipse_sector, gen_rectangle1, gen_rectangle2, gen_circle_contour_xld, gen_cross_contour_xld, gen_ellipse_contour_xld, gen_rectangle2_contour_xld, gen_box_object_model_3d, gen_sphere_object_model_3d, and gen_sphere_object_model_3d_center

• The new operator test_region_points can be used to query which points are contained in one or more regions. Contrary to test_region_point, the test result is returned for each tested point individually. The operator supports parameter broadcasting, with either a single region and multiple points, multiple regions and a single point, or multiple regions and multiple points.

System

• The HALCON third-party library OpenSSL has been updated to OpenSSL 3. Previously, OpenSSL 1.1 was used. The docker example examples/docker/halcon-application/ has been adapted to use Ubuntu 22.04, which supports OpenSSL 3 by default.

  Note that this change affects the compatibility. Read more.

• open_socket_connect now allows setting the "Server Name Indication" (RFC 6066) for TLS sockets using the generic parameter 'tls_sni'.

• get_system_info now supports all parameters that get_system supports.

• set_system and get_system have been extended with the parameter 'enable_neon'. This parameter controls whether NEON operations are used to accelerate selected image processing operators.
  Additionally, get_system has been extended with the parameter 'neon_supported'. This parameter indicates if the processor supports NEON operations.

• send_data and receive_data have been extended with support for the data type 'memory_block'.

• On Windows, HALCON 22.11 added usage of the mimalloc memory allocator (see https://microsoft.github.io/mimalloc for more details) to improve performance in some specific cases. However, this change increased overall memory consumption for all cases. As a result, mimalloc can now be disabled with set_system's new 'memory_allocator' parameter. Additionally, the mimalloc memory allocator is now also available for Linux.

Tools

• distance_pp, distance_ps, distance_pl, distance_sl, distance_ss, angle_ll, angle_lx, and projection_pl have been extended to support parameter broadcasting of the form 1:N and N:1, where each argument can have length N or 1. Additionally, the operators are now automatically parallelized. Finally, the operator distance_pp is now faster by up to 27% for longer input tuples.

• The following operators now support parameter broadcasting, where input tuples of length 1 are automatically extended and used for all created elements: area_intersection_rectangle2, distance_pr, distance_lr, distance_sr, distance_sc, distance_lc, distance_cc_min, distance_cc_min_points, and distance_cc

  Note that this change affects the compatibility. Read more.

Tuple

• HALCON has been extended with the operator tuple_str_replace, which replaces all occurrences of a substring in a string. It is also available as inline operation str_replace.

• HALCON has been extended with the operators tuple_repeat and tuple_repeat_elem:

  • tuple_repeat computes the iterated concatenation of a tuple to itself.
  • tuple_repeat_elem replaces every element in a tuple by a given number of copies of itself.

• The control tuple datatype now supports a copy-on-write mechanism. This increases the processing speed when copying or accessing tuple elements, especially when dealing with long tuples. As a consequence, the processing of dictionaries is accelerated if long tuples are included.

• The access on the key table of HALCON dictionaries is now thread safe. See the reference documentation of create_dict for more details.

• set_dict_tuple can now set multiple keys at the same time. For this, either a single value can be passed, which will be used as value for all keys, or a tuple of values, one for each key. This also applies to dict assignment expressions.

XLD

• The new operator intersection_region_contour_xld can be used to select those parts of a contour XLD that intersect with a region.