Adaptive Part Inspection Through Developmental Vision
We present a novel online inspection method for manufacturing processes that automatically adapts to variations in part and environmental properties. This method is based on a developmental learning architecture comprising a procedure that focuses attention to apparently defective regions, a recognition method that performs automatic feature derivation based on a set of training images and hierarchical classification, and an action step that controls attention and further decision processes. The method adapts to variations incrementally by updating rather than recreating the training information. Also, the method is capable of inspecting and training simultaneously. Addressing new inspection tasks requires neither re-programming and compatibility tests, nor quantitative knowledge about the image set, from a human developer. Instead, automatic or manual training of the inspection system according to simple guidelines is applied. These attributes allow the method to improve online performance with minimal ramp-up time. Our system performed inspection of three applications with low error rate and fast recognition, confirming its suitability for general-purpose, real-time, online inspection.