- Typical defects like hits, scratches or stains are often hardly discernible at the early production stages because of the parts’ rough and strongly textured surfaces.
- Normal variations and non-significant anomalies in the material need to be tolerated.
- Stains & hits often manifest themselves by a local change in contrast caused by non-uniform lighting.
The software algorithm trains itself on a representative set of annotated images as well as known good samples. The learning system automatically incorporates contextual information in order to form a reliable model of the part’s shape and texture. As a consequence, difficult to discern defects can be detected as in the sample to the right: the hits and scratches are considered as anomalies because they have a textured area that deviates from the expected average local surface aspects and tolerable deviations.
See the application notes here.