Labeling Labeling is the process of marking features or defects in images, or categorizing the images into classes. Labeling is crucial for training Deep Learning Lite and Deep Learning Standard tools because the labels serve as training material to guide the tools how to function correctly. is the process of manually marking images for defects or other features of note. During the training process, the neural network model Each AI tool is a neural network model. A neural network model mimics the way biological neurons work in the human brain. The neural network model consists of interconnected layers of artifical neurons, called nodes, and they have multiple layers. Neural network models excel at tasks like image classification and pattern recognition. learns to identify these features with the label applied to the image.

Accurate and consistent labeling is essential for the tool to learn correctly and perform well. The goal of training is to teach the tool to reliably predict the correct class for unseen views based on the labeled training data. OneVision measures the performance of the tool based on how well the predictions match the labels.

Model Types

OneVision provides two different neural network models:

• Deep Learning Lite: this model can train on a smaller image set and produce results quicker, with a slight performance trade-off.

• Deep Learning Standard: this model trains on a larger image set and requires a more time-intensive training period, but produces highly accurate results.

  Note: Deep Learning Standard is currently not deployable on In-Sight vision sensors.

The exact model types vary by AI tool.

Image Sets and Role Assignment

The goal of training a model is to get a model that is good at generalizing on unseen data. To verify your model is working well, you must not test with data that was used for training the mode. Doing so will result in a biased model that reports artificially high results during testing but performs poorly on new or unseen data. To prevent this, users can assign roles to the views used to configuring the tool.

The following role assignments are available:

• Train: Apply this role to views to train the model with. A tool cannot train without views assigned this role. Number of images required varies based on training mode selection.

• Test: Apply this role to views that you would like to use as unseen images to test the trained model with.

• None: Apply this role to images that are not used to train or test the model with. These images are used in the Process step to build the Metrics and Statistics results.

You can manually assign roles to one or more views using the views list panel on the left of the tool configuration screen.

The role assigned to a view are overlayed on the view in the left panel. This overlay is visible on the main display during the train and process steps.