Abstract: Aiming at  the challenge of balancing model accuracy and inference speed in metal surface defect detection, we proposed an improved SWE-YOLOX detection algorithm based on the YOLOX model. Firstly, in order to solve the problem of large  interference of complex backgrounds and significant variation in defect scales, we introduced a channel shuffle attention module to enhance feature expression ability  and suppress irrelevant information. Secondly, aiming at the problem of unclear defect edges and weak texture features, we incorporated wavelet convolution to  improve the extraction ability of frequency-domain features, thereby enhancing the expression of  detailed information. Finally,  the original intersection over union (IoU) loss function was replaced with an enhanced intersection over union (EIoU) loss function to optimize  the regression accuracy between predicted boxes and ground truth boxes. Experimental results show that the proposed method achieves a mean average precision (mAP) of 76.3% 
on the NEU-DET dataset, which is 3.86 percentage point higher than that of YOLOX model. It also maintains a fast inference speed without increasing the number of parameters and  computational complexity.

Key words: YOLOX model, attention module, wavelet convolution, loss function, metal surface defect