A reweighting offset bin classification network for surface defect detection and location of metal components

Abstract The detection and accurate location of metal components surface defects are the basis of automatic laser cladding / laser additive manufacturing repair. Most deep-learning methods utilize regression networks trained with non-maximum suppression and loss function to correct the offset. However, these methods are unable to implement sufficient punishment for accidental offset, resulting in an inaccurate offset prediction. In this investigation, we proposed a reweighting offset bin classification network to solve this issue. The network discretizes the offset values into multiple offset bins by its probability distribution and corrects the offset of the bounding box by predicting the probability of these bins. In addition, the offset bins are classified into different balance groups for parallel training to ensure that all bins have relatively balanced weight norms. Comparative experiments conducted on the advanced detection network and defect datasets demonstrate that our method can locate defects accurately.