Automatic detection of surface defects based on deep random chains

Defect detection is critical in production systems. The traditional methods are primarily manual, prohibiting its large-scale industrial application. The current deep learning methods usually require a large amount of data, which is challenging in some cases. This paper presents a novel deep learning method to detect a large variety of defects based on small datasets only. Specifically, the method is based on the deep random chain combined with the adaptive Faster R-CNN. The idea behind this method is to fuse both common and different types of information among candidate groups to each defect candidate, which thus improves the model's generalization for small sample datasets with a wide variety of defects. Indeed, the deep random chains focus on learning the relationship among the pixels inside each defect, while many features are added to each defect using Faster R-CNN. Several experiments on industrial products demonstrate the merit of the proposed method for small sample datasets with yet a wide variety of defects.