下载中心

MVTEC Deep Learning Tool 25.05 Full Version x64 windows 深度学习工具dlt-25.05完整版

文件名: dlt-25.04.zip
文件大小: 4424550265 字节 (4.12 GB)
修改日期: 2025-04-29 23:48
MD5: b2bfddd937b0220b17ce8b7d32de8f43
SHA1: 6e43c258a82a437d2511190449690665da0fceb5
SHA256: 8b684ec131049c06ae552cdf927dda154833e3beef78da15d3e82b299138178d
CRC32: dd208256

MVTEC官方地址：
需要注册，然后免费下载。www.mvtec.com

百度网盘下载地址：

MVTEC Deep Learning Tool 25.05 Full Version x64 windows 深度学习工具dlt-25.05完整版

http://visionbbs.com/thread-34046-1-1.html

(出处: LabVIEW视觉)

Release notes of previous versions
VERSION 25.04
New licensing and customization features

• New Licensing Mechanism
  

• Resell Capability: Customers can now resell our product as part of their own offerings.

• Non-Expiring Licenses: Professional and OEM licenses will no longer expire, providing continuous access and usage.
  

• Enhanced Customization for OEM Customers
  
  

• Customization Options: OEM customers can now tailor our product to better fit their needs.

• Project Type and Page Restrictions: Supported project types and specific pages can be restricted, allowing for more controlled and secure usage.

• Presetting of Parameters: Parameters for training, evaluation, and export can be predefined, enabled, disabled, or hidden.
  

These updates are designed to provide greater value and flexibility, ensuring our product integrates seamlessly into your solutions.

• The DLT is now based on HALCON 24.11.

• The name of the Hailo WSL container now includes the DLT version, which enables parallel installation of multiple, different DLT versions with Hailo functionality.

• Out of Distribution Detection (OOD) for Classification
  Fitting Out of Distribution Detection (OOD) for Classification has been introduced to the DLT.
  Key Highlights:
  
  

• Enhanced Error Detection: The OOD feature helps recognize unexpected behavior caused by incorrect classifications, allowing users to take targeted measures, such as stopping the machine.

• Deep Learning Integration: When using a deep learning classifier, OOD indicates when an object not included in the training data is classified, providing an “OOD score” to show the degree of deviation from trained classes.

• Model Expansion Support: The OOD score assists in expanding deep learning models by identifying new training images with high information content, enhancing the model's accuracy and robustness.
  

This feature aims to improve production reliability and efficiency by ensuring accurate classification and timely intervention.

• To optimize trained Deep OCR recognition models for a Hailo chip, now a parameterization JSON file is provided.

• The randomization method used for splitting the dataset and during the training has been changed. This improves the comparability of training results with the HALCON scripts, but may result in slightly different training results compared to older DLT versions.

• During the training of a model, the HALCON dictionary file “dl_training_results.hdict” is written to the training folder. Now, this dictionary contains additional entries for each training iteration and validation step. This includes the IDs of the samples that are used for that step and the used iteration seed.

• To visualize the model alphabet, a dialog has been added.

• The documentation now informs about the necessary steps to use adapted postprocessing settings for GC-AD in a HALCON application.

• The evaluation report can now also be generated for Deep OCR projects.

• It can happen that temporary folders created by the DLT are not deleted after a program crash. Now, such folders will be deleted when a new DLT instance is started, as long as the folders are not used by another DLT instance.

• The HALCON Hailo plugin used by the Deep Learning Tool has been updated to version 25.05.5.0, which in turn is based on the Hailo runtime version 4.18. Hence, this is the minimum required Hailo RT version in the inference system now.

• Device precision BF16 can now be selected for supported devices.
  

Resolved issues and improvements

• In some cases, the zoom button of the score histogram did no work. This problem has been fixed.

• If an inference parameter (e.g., device) was changed on the Evaluation page without updating the evaluation, the wrong device was shown in the report. This problem has been fixed.

• When selecting a device with an AI² interface, such as OpenVINO or TensorRT, the Evaluation page shows whether the optimized model is available for that device or not. After resetting the training or continuing it, this information could be wrong: By creating a new best model, all optimized models were deleted, but the display was not updated. This problem has been fixed.

• When evaluating a model optimized for Hailo in the WSL container, an error could occur. This problem has been fixed.

• The Hailo installation did not work if the DLT had been installed on a different drive. This problem has been fixed.

• When training anomaly detection projects, now images from the training split are used for normalization if there are not enough images in the validation split.

• It was possible to set some parameters to invalid floating point numbers. This problem has been fixed.

• On the Evaluation page of a Semantic Segmentation project, the icon for the background class looked different in the confusion matrix and in the match details view on the right side. This problem has been fixed.

• It was possible to enter invalid characters for the parameters “anchor aspect ratios” and “anchor angles”. This problem has been fixed.

• On the Gallery and Review pages, scrolling the images with the mouse wheel sometimes was too slow. This problem has been fixed.

• The documentation did not explain why the F1-score (Deep OCR) or Mean-AP (instance-based projects) shown for the evaluation can differ from the F1-score or Mean-AP shown for the training. This problem has been fixed.

• For Deep OCR projects, exported datasets were not compatible with some HALCON procedures. This problem has been fixed.

• The documentation did not explain the meaning of the advanced network parameter “Only Bounding Box” for instance segmentation projects detailed enough. This problem has been fixed.

• There was a visual bug in the heatmap display when training with non-square neural network image sizes. This problem has been fixed.

• Using augmentation caused an error when the augmentation parameter “Remove Pixel” was used. This problem has been fixed.

• In some situations, the preprocessing of samples was not updated correctly, which could lead to training errors. This problem has been fixed.

• When training Deep OCR detection models on a dataset with upright images, the calculation of the F1-score on the validation and training images after the epoch was wrong. In addition, when showing preprocessed upright images on the Evaluation page, the label boxes were shown at the wrong position. These problems have been fixed.

• When evaluating a Deep OCR recognition model, it is possible to open a Character Overview dialog on the Evaluation page. The first row of the shown character table could contain an empty cell without a specific character. This problem has been fixed.

• In a project with many classes, it could happen that the combo box for setting the label class below the label region was too small to show all label classes. Because there was no scroll bar, the lower label classes could not be assigned to the label. This problem has been fixed.

• On the Training page, the list of augmentation parameters for Deep OCR recognition models was wrong and showed some parameters that cannot be used. Selecting such a parameter caused the training to fail. Vice versa, the valid parameter “Rotate Range” was not provided. In addition, activating just the parameters “Remove Pixel X/Y” caused the display of a warning that no option had been selected. These problems have been fixed. For new trainings, images can now be augmented with Rotate Range value between 1 and 5.

• On the Image page, it could happen that the popup for displaying and changing the class or text of a label was shown too close to the right border, thereby leaving no space for the “Delete Label” button. This problem has been fixed.

• The Deep Learning Tool could crash if the number of epochs was set to 0 and then a training was started. This problem has been fixed.

• If the split of a training had no training images, pressing the Edit / Preview button on the Training page could crash the Deep Learning Tool. This problem has been fixed.

• On the Image page, some parameters may have appeared hidden. To give a hint that there are further parameters, now scrollbars are available.

• If a label class of a project was changed after a training was performed (by renaming, adding, or deleting a class) the evaluation of the trained model could become corrupted and the assignment of all predicted classes could be completely wrong. This problem has been fixed. If the model now predicts a class that is no longer defined in the project, that prediction is visualized as undefined class. All other class predictions are handled correctly.