Efficient and Accurate Detector with Global Feature Aggregation for Steel Surface Defect Detection

Kefei Qian, Zhiwen Wang, Lai Zou

Abstract: Object detection-based steel surface defect detection is a typical application of deep learning technology in the industry, which performs better than traditional detection methods. However, simply using popular networks and porting high-performance modules does not yield the desired effect on steel due to the vastly different datasets. In response to the problems of large-scale variation, inconspicuous features, and substantial background interference of steel surface defects, we propose a lightweight global feature aggregation and redistribution (GFAR) module that significantly improves detection accuracy while ensuring efficiency. Our method uses yolov5m as the baseline network and embeds the proposed module at the neck: adaptive aggregation of feature maps in a parameter-free manner, output in the same and opposite manner after feature refinement and refusion with the feature maps of the PAFPN branch. The proposed network is extensively experimentally validated on the steel surface defect datasets NEU-DET and GC10-DET, and the results show that our method achieves a relatively competitive level of accuracy while ensuring the detection speed.