Improved YOLOX-based Method for Steel Surface Defect Detection

Surface defect detection plays a crucial role in the steel production process, but most current detection methods are inefficient and expensive. However, with the advancements in deep learning–based object detection technology, an improved balance has been achieved between detection accuracy and speed, allowing for enhanced learning of complex defect features and meeting real-time detection requirements. In this study, we present an improved YOLOX-based approach for detecting steel surface defects. Our method incorporates attention modules in the backbone architecture to improve the extraction of defect features. Furthermore, we propose a channel product module that uses dot product operations between feature channels to enhance the feature expression ability of the model through information interaction and fusion across different channels. The results of our approach on a steel surface defect dataset (NEU-DET) and a steel rail dataset prepared herein indicate that the proposed approach achieved a mean average precision of 76% and 88.1%, respectively. The detection speed exceeded 80 frames per second (FPS) for datasets, demonstrating high detection accuracy and speed. These results highlight the practical value and effectiveness of our method for steel surface defect detection.