Release Notes

Mech-DLK 2.4.1 has added a brand-new cascade mode of modules, which enables the combination of modules to solve deep learning problems in complex scenarios. It should be noted that when the Fast Positioning module is involved, it must be the first module. For example, when you need to detect the positions of defects and classify these defects, you can first add a Defect Segmentation module and then add a Classification module. In addition, when importing data from the previous module, you can select images according to actual needs and configure the import of these images in the Import window.

The Training Center supports the training of models in a queue, which is suitable for scenarios requiring the training of multiple models. With the Training Center, the software can train models in sequence, with no need for manual clicking of Train repeatedly, which can save a huge amount of time.

In the Defect Segmentation module, when you select the Mask Polygon Tool, the mask type can be Single Image Mask or Global Mask. In addition, you can customize the mask color.

Click in the lower-right corner of the selection region to open the window of keyboard shortcuts.

In the Instance Segmentation and Object Detection modules, auxiliary labeling lines are added for the Rectangle Tool to assist rectangular selection on images.

In the Instance Segmentation and Object Detection modules, a confidence filtering function is added for validation results. You can adjust the confidence to filter the validation results and then evaluate the accuracy of models.

The Classification module is optimized, which leads to faster training convergence and a growth rate of 20% in accuracy under complex scenarios.

Mech-DLK SDK is more stable and easier to use after being restructured. Mech-DLK SDK supports the inference based on the cascaded modules and the switching between different operating hardware. In addition, it provides richer samples for reference.

The defect determination rule in the Defect Segmentation module is optimized. Click here to view the details.

In the Image Adjustment window of the Fast Positioning module, you can translate the image along the X and Y axes. After training, images with objects in specified positions and orientations will be generated, which meet the requirements of more application scenarios.

In the Instance Segmentation and Object Detection modules, after selecting the Template Tool, press and hold the Shift key and scroll the mouse wheel to adjust the angle of the template. You can also set the Rotation angle to achieve the same purpose.

Before using Mech-DLK 2.3.0, please upgrade the graphics card driver to 472.50 or above.

Optimized the algorithms, and thus significantly improved the speed of model training. Only the optimal model is saved during training, and the training cannot be stopped halfway.

For modules including Defect Segmentation, Instance Segmentation, and Object Detection, you can do smart labeling by selecting the Smart Labeling Tool, clicking the objects to be labeled, right-clicking to undo the redundant selection, and pressing the Enter key to complete the labeling.

For the Instance Segmentation and Object Detection modules, after labeling with the Polygon Tool, if the selection needs to be modified, you can left-click the line segment between two vertices to add a vertex, or right-click a vertex to remove it.

For the Instance Segmentation and Object Detection modules, you can use the Template Tool to set the selection as a template. The template can be applied by simply clicking the images. It is suitable for scenarios where there are multiple neatly arranged objects of the same type in an image, and it improves labeling efficiency.

Support previewing full images and cropped cell images.

Optimized the Grid Cutting Tool. After cutting the image by the grid, you can select a cell image by checking the box in the upper left corner of the cell image, and you can preview the image by clicking on the button in the upper right corner of the cell.

Added options for filtering results: “Correct results”, “Wrong results”, “False negative”, and “False positive”. Added options for filtering data types: “Labeled as OK” and “Labeled as NG”.

The deep learning environment is built into the software Mech-DLK, and the models can be trained without a separately installed environment.

After the model is trained, click Generate CAM. The class activation maps show the weights of the features in the form of heat maps; the model classifies an image into its class according to these features. Image regions with warmer colors have higher weights for classifying the image into its class.