System concept for image sequence classification in laser welding

Abstract

This work introduces a concept for a challenging classification problem of highly dynamic laser welding image sequences. Under the conditions of the industrial production a decision is required within seconds with a high selectivity. The introduced concept consists of classic elements of image processing, classification, and machine learning. The components are not optimized but as a whole, yielding a high-performance system. It also contains individually a statistical preprocessing for change detection in image sequences. The event hypotheses generated by the algorithm are classified frame-wise by an object classifier. The optimization of recognition performance is based on a feature selection by a modified sequential forward selection. We use the properties of the polynomial classifier for the efficient computation of cross validations. © Springer-Verlag Berlin Heidelberg 2003.