This new HALCON feature makes it easy to recognize unexpected behavior caused by incorrect classifications in production. Thus, users can take appropriate measures, such as stopping the machine, in a targeted and efficient manner. When using a deep learning classifier, unknown objects are assigned to one of the classes that the system has learned. This can lead to problems if, for example, the defects or objects themselves are of a type that has never occurred before. The new deep learning feature “Out of Distribution Detection (OOD)” indicates when an object is classified that was not included in the training data. For example, this could be a bottle with a green label if the system was only trained on bottles with red or yellow labels. In such cases, HALCON provides the message “Out of Distribution” together with an OOD score that indicates how much the deviation from the trained classes is.

The OOD score can also be useful when expanding deep learning models with new training images by indicating which of the new images will have the greatest value for the new model. For example, a high OOD score for a new training image indicates a greater deviation from the images already in the network – this means a higher information content and, therefore, greater value for the training.