Self-Organizing Map Based User Interface for Visual Surface Inspection

In visual surface inspection applications, a problem often faced is the collection and labelling of training material. The manual labelling of training samples requires much effort and is error prone since the defect classes may appear difficult to discriminate even for a human. Frequent changes in the inspected material or imaging conditions may lead to impractical often repeating training material collection cycles. We propose a SOM (Self-Organizing Map) based classifier and user interface scheme for visual surface inspection problems. The approach combines the advantages of non-supervised and supervised classification. The SOM based approach supports the labelling of training data, simplifies the retraining for changing material or imaging conditions, provides an intuitive user interface, and is computationally attractive for real-time applications.