Physically-based data augmentation for deep learning-enabled automated visual inspection of scratches

This paper studies the problem of surface defect detection of metal parts in small samples. In the production process of some important metal parts, such as surface defects of aircraft engine blades, it is usually difficult to obtain large quantities of surface defects of these metal parts, resulting in relatively few defect sample data. However, automatic detection of surface defects in metal parts based on deep learning requires a large number of training samples as data sets during the training process to achieve good results. In order to achieve this goal, and in view of the problem of insufficient surface defect data sets of important metal parts, we constructed a physical simulation synthetic metal surface defect generation model to expand the surface defect data sets and improve the recognition accuracy. Moreover, we constructed a semantic segmentation network model suitable for surface defect detection in this study, which is a basic model for detecting surface defects. In addition, experiments have proven that our method can improve the detection accuracy of metal surface defects.