Steel Defect Detection Based on Modified RetinaNet

Abstract

The detection of possible defects on the steel surface is the last step and considered as the last line of defense for manufacturers to ensure the quality of products. To identify defects on the steel surface more quickly and accurately, this paper proposes a modified network based on RetinaNet in view of the current difficulties in detecting large-area steel defects and the overall performance to be improved of detectors. The depth of the FPN is increased to improve the ability to extract semantic information; the detail information enhancement structure, the feature information optimization structure, and the semantic information enhancement structure are combine into the augmented feature information to make the network better in using of image information; finally, we increase the sub-network’s layers to improve the ability to classify. Experiments were carried out on DGAM 2007 and PASCAL VOC. DGAM 2007 is a steel defect dataset, and PASCAL VOC is a general dataset to verify the overall performance of the detector. The results show that our proposed network has greatly improved the detection accuracy of the steel surface defects to 99.73%, and to some extent, large-area defects can be easily detected, and the mAP on DGAM 2007 and PASCAL VOC is both higher than that of the original RetinaNet network, which proves that the modified algorithm can play as an outstanding detector.