下载中心

MVTEC Deep Learning Tool 25.12 Full Version x64 windows 深度学习工具dlt-25.12完整版

文件名: dlt-25.04.2.zip
文件大小: 4426309811 字节 (4.12 GB)
修改日期: 2025-09-26 22:25
MD5: f44fb61a6fbcf152286bee9ac1919e1c
SHA1: bad970adc972778c6ce91e734b83c67ef3e66b6c
SHA256: 073fee5ad30f9cdf163f80e7234f40f0797f50cd9afdcca38bceaa59605371c0
CRC32: c5f80aa6

MVTEC官方地址和百度网盘下载地址：

MVTEC Deep Learning Tool 25.12 Full Version x64 windows 深度学习工具dlt-25.12完整版

http://visionbbs.com/thread-36125-1-1.html

(出处: LabVIEW视觉)

MVTec Deep Learning Tool 深度学习工具 DLT25.12 现已发布
2025 年 12 月 04 日，我们发布了 MVTec 深度学习工具的新版本 25.12，从而进一步发展了我们基于人工智能的机器视觉产品组合。
新功能
图像管理的标签

标签是一种灵活的工具，为深度学习工具中的数据管理提供了许多可能性。标签是一个与一个或多个图像相关联的用户定义字符串。其用途与深度学习工具的众多应用一样多样。

加载和训练早期版本的 HALCON 模型

使用深度学习工具专业版，用户现在可以加载早期 HALCON 版本的预训练模型。这使得可以持续维护和重新训练在新版深度学习工具中被替换的模型。

在硬盘上移动和重命名图像

用户现在可以在深度学习工具中移动和重命名图像。无需再切换应用程序来移动或重命名。这也会更改您硬盘或文件共享中的图像。

编辑验证率以加快训练速度

用户现在可以在训练过程中降低甚至禁用验证，以加快训练速度。

变更日志

Version 25.12

New Features新功能

• 
  It is now possible to create tags and assign them to images. An image tag is an arbitrary text, and it can be used as a filter criterion.

• 
  It is now possible to double-click the opacity, brightness, and contrast sliders to toggle between the new and the default value.

• 
  After renaming, adding, or deleting label classes in a project with trained models, the label classes of the model and the project may differ. If an evaluation is performed in such a case, the automatic evaluation cannot be complete because the mapping between predicted classes and the ground truth is lost. This information is now shown in the confusion matrix and in other evaluation details, such as the class overview, evaluation item tooltips, and the evaluation item details list.

• 
  The Training page now offers a new dialog that allows loading external DL models for training (only available for Professional and OEM edition).

• 
  The light theme has been improved.

• 
  The fonts used for the documentation pages now are consistent with those used in the UI.

• 
  The look and feel of the evaluation report has been improved.

• 
  On the Training page, the list of pretrained DL models now reflects the model files that are actually available. I.e., now it is possible to remove single model files that are not needed from the dl folder, so that they will not be offered in the model selection combo box.

• 
  It is now possible to adjust the frequency of the validation by configuring a parameter of the Advanced Training Parameters.

• 
  The documentation now provides more information about measuring inference times and gives hints on how to achieve comparable results with HDevelop.

• 
  It is now possible to jump to a specific image index on the Gallery and Image pages using the keyboard shortcut Ctrl+G.

• 
  It is now possible to move and rename images from the Gallery and Image pages.

• 
  It is now possible to navigate to a specific image index on the Gallery page.

• 
  When creating a split, it is now possible to consider only the images from the active filter result.

• 
  When switching between images, the currently zoomed-in area is now preserved instead of the new image being fitted to the screen.
  

Resolved Issues and Improvements解决的问题和改进

• 
  In the OEM version of the Deep Learning Tool, setting the batch size for the Evaluation page via a JSON configuration file was broken. This problem has been fixed.

• 
  In the OEM version of the Deep Learning Tool, hiding the random seed parameter on the Training page via a JSON configuration file did not work as expected. This problem has been fixed.

• 
  Training a model for semantic segmentation failed with an error message if the dataset was imported from an HDICT file in which the background class was set to a class with a class ID other than 0. This problem has been fixed.

• 
  In Deep OCR projects, there were two problems with the int8 model optimization. When the model size was changed on the Evaluation page, the optimization failed, and the performance of an optimized recognition model was very poor. These problems have been fixed.

• 
  The documentation did not mention the “Guided Grad-CAM” feature, which is available for Classification projects. This problem has been fixed.

• 
  When creating a split, in some cases it was not possible to fully delete the proposed split name. This problem has been fixed.

• 
  Using HTML code in class names or labels could result in a corrupted display. This problem has been fixed. Now, HTML is treated as plain text rather than being rendered.

• 
  The display names of the pretrained models that are used for Deep OCR have been adapted to be closer to the names used in HALCON. The recognition model is now shown as “OCR Recognition Default” and the detection models as “OCR Detection Default” and “OCR Detection Compact”.

• 
  In some cases, the IoU threshold used for evaluating a model was not correctly transferred to the generated report. This problem has been fixed.

• 
  The description of the Deep OCR example project now mentions that some images intentionally are unlabeled to be able to demonstrate the auto-labelling feature.

• 
  The action "Show label / image on review page" could lead to an unexpected result if the corresponding label class was filtered out. This problem has been fixed.

• 
  If a project without trainings was saved under a different name, overwriting an existing project with the same name that contained trainings, the old trainings were not deleted. This problem has been fixed.

• 
  The import path of a procedure in the anomaly_detection_global_context_inference.hdev example did not resolve. This problem has been fixed.

• 
  When creating a custom image filter, in some cases it was not possible to fully delete the proposed filter name. This problem has been fixed.

• 
  The documentation of the manual Hailo WSL install commands has been improved.

• 
  If a model was trained in a project and the ground truth was later changed (for example, by changing or deleting label classes or updating a text box or the text) subsequent trainings could fail or use outdated data. These problems have been fixed.

• 
  The Character Overview table, which can be opened on the Evaluation page when evaluating an OCR recognition model, was broken. This problem has been fixed.

• 
  For newly created Deep OCR projects, the setting for automatic text recognition when drawing regions on an image was not evaluated, which could lead to unexpected behavior. This problem has been fixed.

• 
  If a classification project was created from an HDICT dataset file and a label class was removed, the training could fail with a message that the HALCON class IDs were not consecutive. This problem has been fixed. The same problem could occur if the class IDs in the imported dataset file were not consecutive. Now, in these cases, the HALCON class IDs are reassigned before training so that they are consecutive.

• 
  When importing a dataset without any samples, the dataset import dialog stated “First image found” although there was none. This problem has been fixed.

• 
  The default value of the evaluation parameter “evaluation_interval_epochs” was not set. This could lead to false behavior if the parameter was missing in the model. This problem has been fixed.

• 
  The “+” button in the split creation dialog was not disabled when the maximum number of images had been reached. This problem has been fixed.

• 
  In Instance Segmentation projects, the setting “Instance Segmentation: Only bounding box” was not evaluated correctly, resulting in false evaluations. This problem has been fixed.

• 
  In Deep OCR projects, a required optimization step was not automatically initiated when an evaluation was started. This problem has been fixed. Now, the necessary optimization runs seamlessly as part of the evaluation process.

• 
  In some cases, the DLT could not load large images. This problem has been fixed.

• 
  If no suitable device was available, there was a problem with the selectable device for training. This problem has been fixed.

• 
  For Hailo devices, the option to use an alloc script and the option to finetune were not displayed on the Evaluation page. Further, a displayed warning message was not reset under some circumstances when changing the DL device. These problems have been fixed.

• 
  There was an issue leading to an unreadable configuration file in the example directory. This problem has been fixed.

• 
  The selected split for the training was not considered on the Evaluation page. This problem has been fixed.

• 
  The documentation missed explaining the “Limit samples per class” parameter for Out-of-Distribution detection. This problem has been fixed.

• 
  When fitting for out-of-distribution detection, the state was falsely displayed for all trainings. This problem has been fixed.

• 
  Large confusion matrices were sometimes not fully visible. This problem has been fixed.

• 
  The command line option --use-rtl has been removed.

• 
  After the renaming of a training, it could happen that the training could not be used anymore without closing and reopening the project. This problem has been fixed.

• 
  After resetting a training that already had been evaluated, the evaluation result shown in the training list on the Training page was not cleared. This problem has been fixed.

• 
  After a training was stopped, the pause and stop buttons did not show the correct state. This problem has been fixed.

• 
  In the new project dialog, the project path is checked for validity. This check was too restrictive, so some valid paths with spaces before or after a dot could not be used. This problem has been fixed.

• 
  Setting a non-existing model in the auto label configuration dialog for Deep OCR projects could lead to a shutdown of the Deep Learning Tool. This problem has been fixed.

• 
  Paused evaluations for OCR Recognition were reset when resuming. This problem has been fixed.

• 
  When trying to close a project via the Quit menu entry or the Ctrl+Q shortcut and then clicking Cancel, a second instance of the dialog appeared, requiring another click of the Cancel button. This problem has been fixed.

• 
  On a system without GPU, it could happen that the device for automatic text recognition was set to “gpu” even if no GPU device is available. This problem has been fixed.

• 
  The alphabet dialog could not be opened after a training was started. Thus it was impossible to view the alphabet used for training. This problem has been fixed.

• 
  It is now possible to change the parameters in the Preprocessed Image Preview dialog by directly editing the values in the text boxes.

• 
  The documentation did not mention how to start the DLT using a configuration or theme file. This problem has been fixed.

• 
  The fonts used for the texts of the quick help were different from those of the DLT program itself. This problem has been fixed.

• 
  In some cases, the displayed evaluation results could differ between Training page and Evaluation page. This problem has been fixed.

• 
  The runtime of a trained model was not correctly set in some cases. This problem has been fixed.

• 
  The model fitted for Out-of-Distribution detection was not exported, even though the option was selected. This problem has been fixed.

• 
  In Instance Segmentation projects, the value set for Mask Heads Weight was not saved. This problem has been fixed.

• 
  In anomaly detection projects, the postprocessing parameters were not written into the exported model. This problem has been fixed.