Surface Defect Detection of Steel Based on Improved YOLOv4

Abstract

An important measure of the nation's industrial level is the development of the steel sector. And the surface quality of steel is related to the development of high technology such as automotive, energy and avionics. To address the issues of traditional steel surface flaw detection's low accuracy and slow speed, a steel surface defect detection method based on large convolution kernel module combined with channel segmentation and re-parameterizing mechanism is designed. A large convolutional kernel is introduced into the YOLOV4 backbone network to increase the effective receptive field of the feature layer and improve the feature extraction capability of the network. At the same time, a channel segmentation technique is introduced to enable the network to learn differentially as much as possible to improve the generalisation capability of the network. To increase the detection accuracy of the approach for a variety of complicated flaws, a tiny convolutional kernel parallel to the large convolutional kernel is employed to re-parameterize the network structure during model training. According to the testing findings, the enhanced algorithm's average accuracy mean mAP is 91.15%, which is 3.55% better than the baseline YOLOV4 algorithm, and its single picture detection speed FPS is 6.48 f/s, which is sufficient for the identification of steel surface defects.