A Learning-Based Surface Defect Detection Method for Internal Surfaces of Bellows

Recently, vision-based defect detection methods have been proved effective and found many applications. In this paper, we investigate the detection of defects of internal surfaces of bellows. However, the existing defect detection methods cannot be simply implemented to accomplish this defect detection task, since the defects here scatter in the bellows and may have quite different appearance. To this end, we propose a novel defect detection method with two fundamental vision branches which target at classification and object detection, respectively, and can greatly boost the defect detection performance. The classifier aims to find the defects with obvious appearance while the detector is designed to find small defects that are more difficult to recognize. To improve the accuracy of the detection branch, classification feature maps are integrated into the detection framework to enrich the target information. Our method was verified through some bellow images and achieved an accuracy of more than 75 %.