Steel Surface Defect Detection Method Based on Improved YOLOX

Steel is a crucial material that is extensively utilized in various aspects of daily life and holds significant importance. However, during its production, there is a possibility of certain defects arising that could have a negative impact on the quality of the steel. Only by accurately detecting the defects on the steel surface can we avoid the harm caused by defects in steel. Because the steel surface defect detection algorithm is prone to misdetection, missed detection, and other problems, a steel surface defect detection algorithm based on improved YOLOX is proposed. First, the CSPCrossLayer module proposed in this paper is used to replace the CSPLayer structure in the backbone network to enrich the gradient information of the network and strengthen the feature extraction capability; Then, the SA (Shuffle Attention) module is added after the output of the backbone network, highlighting the general information to input high-quality features for the feature fusion network; Finally, the PSblock module is proposed to replace the CSPLayer structure in the feature fusion network, which reduces redundant computations to efficiently perform feature fusion on feature layers of different scales and improves the feature fusion capability of the model. The experiments involved testing the algorithm on two datasets: the publicly available NEU-DET dataset and a steel rail dataset collected in this paper. The algorithm can reach 77% mAP on the NEU-DET dataset, while the detection speed is 100 FPS. It reaches 88.8% mAP on the steel rail dataset, and the detection speed is 93FPS. These results demonstrate that the proposed algorithm is capable of swiftly and accurately detecting surface defects in steel materials.