Steel Surface Defect Detection Method and Application Based on Improved CenterNet

Defect detection is vital for steel production. Manual detection and classical algorithms are the primary detection methods. Manual detection is more thorough, reliable, and dynamic than other detection methods but requires workers to focus for extended periods on high-intensity activity, which is harmful to their health. Although the typical detection algorithm is very mature, the features must be designed manually. It has strict professional background requirements, low generalization, and is strongly impacted by environmental conditions such as illumination and shadow, limiting its application breadth. To address this issue, a steel surface defect detection method based on CenterNet was proposed in this study. By enhancing Resnet50, we improved the feature extraction ability of this network, and the detection accuracy was significantly improved. At the same time, the algorithm could provide an attention module to emphasize valuable traits, while ignoring unnecessary ones. The results revealed that the algorithm's average accuracy on NEU-DET was 74.5%, and the detection speed was 123.4 frames per second, indicating a reasonable balance between detection precision and detection speed, which could meet the needs of the industry.