Segmentation Block

Use the Segmentation Block page to train a deep-learning model to take an image and identify each pixel's content.

Capture — This lets you take a new Capture and use it to train a model.

Import From Library — This lets you add images stored in your Library for training.

View All ROIs — This modal shows each labeled ROI cropped from its parent image. It lets you easily visualize all of your training data and verify that they are all labeled correctly and consistently.

Tips to check label integrity

Set the opacity slider to 0 to see the images without any labelling overlays and then increase it slowly to 0.5 to make sure the overlays are drawn precisely.

< Prev — This lets you view the previous captured image.

> Next — This lets you view the next captured image.

Tips to Navigate across captures

You can enter the capture number to navigate to a particular image directly.

Train Segmentation Model — This lets you train a segmentation model.

Live Preview Mode — Runs the alignment, ROI, and segmenter blocks in real-time allowing you to test its performance and confidence on different situations and parts.

Confidence Score Threshold — This lets you control the model’s sensitivity to detecting defects. If you want the model to be very sensitive and detect more things, decrease the slider. If you want the model to be more selective in detecting things, increase the slider.

Note

You can only adjust this setting when Live Preview Mode is enabled.

Labelling Data

Pan — This lets you pan/move the image around.

Brush — This lets you draw on the image. You can use this to highlight the areas of the image that you want the model to learn to detect.

• Use the Brush Class dropdown to select the class of the area you are going to be highlighting.

• Pick a Brush Size that lets you precisely cover the part of the image that you want the model to learn to detect.

Eraser — The eraser lets you erase the annotations that have been drawn.

Show Annotations — Toggling this off lets you see the image without the user drawings overlayed on them and toggling this on lets you see the image with the user drawings overlayed on them.

Drawing on the image is not sufficient to add the ROI to the training dataset. You have to explicitly check the checkbox present at the top left of the ROI to add it to your training dataset.

Segmentation Labelling Best Practices

• The segmenter model will roughly mirror the precision of how well you paint/label the training images. So, try to be as precise as possible when labelling them.

• In your efforts to draw precisely, if you have to decide between slightly overdrawing and fully covering the defect or slightly underdrawing and leaving out a part of the defect, choose to overdraw and fully cover the defect.

• When annotating a segmenter training image, you are showing the model what to detect by drawing on it, but you are also showing the model what to ignore through all the parts of the image where you have not drawn. So, make sure to draw on all instances of the defects you are trying to detect in the image.

• Adjust your imaging setup to make sure that each instance of the defects that you are trying to detect are at least 6 pixels in radius (you can use the brush sizes for reference) to give you the best odds at training a successful model. You can get a production grade model even on defects that are less than 6 pixels in radius. This is just a good setup to aim for.