YOLO-PMSW: An Improved Algorithm for Efficient and Lightweight Steel Surface Defect Detection

In this paper, an improved lightweight inspection model, YOLO-Pyramid Multi-Scale and Shared-Weight Network (YOLO-PMSW), is proposed to address the problems of leakage, false detection and high model computational complexity in steel surface defect detection. The model optimizes the performance through the following innovative strategies: first, the C2f-Multi-Scale Residual Dilated Fusion (C2f-MRDF) module is designed to enhance the multi-scale feature extraction capability and improve the adaptability to the fine cracks and complex background defects. Second, the SPPF-Pyramid Dilated Convolution Module (SPPF-PDCM) module is proposed to optimize the model structure by combining the expansion convolution to efficiently extract detail and global information. Again, the Detect-Shared Multi-Scale Lightweight Detection Head (Detect-Smld) module is constructed to realize multi-scale feature fusion through shared convolutional kernel, and the task decoupling design is introduced to improve the classification and regression performance in complex scenes. Finally, the Wise Efficient Intersection over Union (Wise-EioU) loss function is introduced to optimize the target frame regression strategy and enhance the robustness of the model to different targets. The experimental results show that the improved algorithm has significant improvement in detection accuracy, model lightweight and real-time performance, and can effectively meet the demand for high accuracy and efficiency in industrial production.