HALCON 25.11: Increased Speed for Deep Learning Applications

In the new version 25.11 of HALCON, which will be released on November 12, 2025, the software will be expanded with additional deep learning features. The focus is on speed. “Speed is a crucial factor in ensuring that deep-learning-based applications can be used profitably in practice. For the new HALCON version, we have optimized the deep learning models for classification and code reading. As a result, inference is accelerated many times over,” says Jan Gärtner, Product Manager HALCON at MVTec. Speed is also the focus of the new Continual Learning feature, particularly when adapting to changes in production processes. The new deep learning model for classification tasks makes it possible to quickly and easily add both new classes and new image data for retraining – without the effect of catastrophic forgetting occurring.
As with every new HALCON version, which is released semiannually, HALCON 25.11 includes not only new features but also improvements to existing ones. In addition to new developments and optimizations, MVTec continuously incorporates customer feedback and integrates corresponding requests into HALCON. One example in the new version is improved visualization for matching methods: both detected and undetected edges are now displayed in different colors.

Continual Learning – Classification

HALCON 25.11 introduces Continual Learning – Classification, enabling faster, flexible training and updating of classification models with few images per class. Existing classes can be refined or new ones added anytime, without full retraining. The approach avoids catastrophic forgetting and runs even on edge devices, eliminating external hardware. This provides a flexible solution that adapts to changing production conditions and suits embedded and edge environments such as smart cameras and sensors.

Score Visualization for Shape Matching

With HALCON 25.11, Score Visualization for Shape Matching increases transparency when setting up applications. Instead of only returning an overall score, it shows how different model contours contribute. Color-coded bins highlight which areas match well and which couldn’t be found, e.g., due to shadows or unwanted textures. This feedback helps refine models and optimize applications. The feature also supports advanced robotics scenarios, such as analyzing which object in a stack is least covered.

Optimized Deep OCR models for faster, resource-efficient OCR

HALCON 25.11 introduces new Deep OCR recognition models for faster, resource-efficient text reading without accuracy loss. They deliver up to 50× faster inferences on embedded devices. Pretrained on industrial data and including proven alignment preprocessing, they enable real-time OCR even on low-power hardware. This makes them ideal for demanding inline tasks such as serial number inspection, label verification, and lot tracking.

MobileNetV4 classification models

With HALCON 25.11, MVTec adds support for the MobileNetV4 series – efficient deep learning models optimized for resource-constrained systems and edge devices. They enable both classification and object detection with high accuracy at low computational cost. Users benefit from faster inference, lower system costs, and simple integration into existing projects. All models are pretrained on industrial data and deliver strong results for tasks such as quality inspection, product classification, and defect analysis.

Various code reading and print quality inspection improvements

HALCON 25.11 makes code reading and print quality inspection (PQI) more robust. QR code detection was improved for difficult cases like curved or deformed surfaces, with faster runtime in standard scenarios. The bar code reader is more tolerant of irregular bar widths in Code 128 and GS1-128. HALCON also supports the latest PQI standards ISO/IEC 15415:2024 and ISO/IEC 29158:2025, ensuring up-to-date compliance in industries such as logistics, food, and pharma.

Built-in SBOMs for easier compliance

HALCON 25.11 delivers Software Bills of Materials (SBOMs), giving users transparency into included software components. SBOMs are increasingly required under regulations such as the EU Cyber Resilience Act. Provided as SPDX JSON files, they simplify compliance, support vulnerability and license checks, and reduce effort and long-term cost.

Legal Notice
All rights reserved, Copyright ©MVTec Software GmbH. All image and text material may be used solely for press coverage for free. Any kind of commercial usage of the material beside press coverage is prohibited.

We would be pleased to receive a voucher copy.