Search

Precise locating and correct classification of each defect in the image is the primary goal of surface defect detection (SDD). This is very similar to object detection, which seeks to predict a set of bounding boxes (Bboxes) and category labels for each object of interest. However, due to the fact that the category dependencies among defects are obviously different from those among general objects, the methods developed for object detection cannot be directly utilized for detecting defects. To address this issue, we proposed a category-related non-maximum suppression (CR-NMS) method. Different from most NMS methods, CR-NMS use Cover Percent (CoP), instead of Intersection over Union (IoU), to guide the Bbox regression. Moreover, a two-stage Bbox regression algorithm is proposed to remove the duplicate Bboxes. This algorithm works in a course-to-fine manner and takes the correlation between Bboxes into account. By this means, the CR-NMS can remove the duplicate Bboxes more effectively and is easy to embed in existing neural networks. A wafer surface defect dataset including 6,000 images and 11 defect categories was set up for training and evaluating our method. Experiments demonstrated that in the detectors, which were built on Faster R-CNN and RetinaNet with backbone ResNet-101, and YOLOX with backbone CSPDarknet, compared with STD NMS or Soft NMS, CR-NMS increased by 3.4%, 9.3%, and 5.5% on mean average accuracy (mAA), and their values were as high as 92.3%, 91.8%, and 90.8%, respectively. The same tests were also performed on dataset NEU-DET. The results showed that compared with STD-NMS and Soft NMS, the CR-NMS can also obviously improve the performance of all the detectors, and their mAA was improved by at least 1.2%. [ABSTRACT FROM AUTHOR]