You can use the classifier block to categorize each inspection region from the ROI block into different categories, perform presence-absence checks, defect detection, assembly verification, etc,

☑️How To:

  1. Click on the Edit option under the Classifier Block


  2. For each Inspection Type you are creating under each recipe, repeat the following process until you have set up all classes:


  3. Select Edit next to the Inspection Type and a new class will be generated

    1. Modify to represent the classification type that you are setting by clicking on the pencil icon next to the Inspection Type you are editing.
    2. Enter the class name and select🆗
  4. Select the Capture option from the main page, and then select capture again once the live image has been activated


  5. Click on +Add New Class to generate all of the chosen class designations for each Inspection Type until completed**:**

    1. Creating a new Inspection Type will automatically add a default Good and Bad class.
    2. Select the color option to control how the designations will be represented


  6. Select the class type e.g., Good under Letter Verification and click on the respective ROIs to set the appropriate class. To change the ROI to “Bad” —> use your cursor to double click within the selected ROI and this will change the designation


  7. Once you have at least five representations for each classes for each Inspection Type, click on the Train Classification Model and the system will generate your fully trained AI model.


💡Tips & Best Practices

Dataset Collection

  1. You should ensure that the features that differentiate the inspection region between multiple classes are clearly visible on the image to be able to train a performant classifier.

  2. To create a highly performant classifier, you should collect data with these principles in mind.

    1. Diverse - You should collect images under conditions that represent all the different variations you expect to see in production. Changing ambient lighting, placing objects in different places in the field of view, etc.

    2. Label Consistently - You should ensure that all images are labeled correctly and consistently. You can check this through the View All ROIs functionality.

    3. Large in Number - The more images (diverse) you provide the classifier to learn from, the better it can learn.


      Example of a good setup with consistent labelling and diverse images.

  3. You should try and create as balanced a dataset as possible. Avoid situations where you have > 10x images in one class than the other.


    Example of a Balanced Dataset


  1. Number of Iterations - How many times we show the labelled images to the model to help it learn. We recommend incrementing in units of a 100. If you see that model accuracy slowly increased even towards the end of training, you should retrain with more iterations.
  2. Finish Training Early - If the model has been at 100% accuracy for at least 20 consecutive iterations, you can consider using Finish Training Early to stop the training process.

What to do when the reported training accuracy is low ?

  1. If the training accuracy is low, check the image labelling through View All ROIs to ensure that all images have been labelled correctly and consistently.