ST-YOLO: An Improved Metal Defect Detection Model Based on Yolov5

In industrial applications, surface defects in metallic materials are inevitable, but they can have a great impact on industrial reliability. Moreover, the detection of metal surface defects is also a challenge due to excessive risks caused by natural corrosion, manual maintenance costs, and harsh environments. Traditional detection machine learning and visual detection methods have the shortcomings of low accuracy and poor effect. Due to limitations, our study introduces Swin Transformer and Wise-IoU Loss methods. Swin Transformer leverages the pooling technique of convolutional neural networks to facilitate patch merging, merging effect improves the model's ability to capture multi-level features in images. Wise-IoU introduces an attention loss function based on bounding box regression (BBR). The design weights of this function exhibit both monotonic and dynamic changes, allowing for dynamic adjustment based on real-time training conditions. Experimental results conducted on NEU-DET and GC10-DET data sets demonstrate the effectiveness and superiority of the model. By using the [email protected] metric, the model's accuracy reaches 84.73% and 67.85%. In comparative experiments, our model outperforms other control models. These results are significant improvements of 9.6% and 2.04% over the baseline Yolov5s model. Furthermore, the model played a positive case in 5 out of 6 scenarios in the ablation experiments.