MPA-YOLO: Steel surface defect detection based on improved YOLOv8 framework

In response to the growing demand for high-quality steel, the detection of surface defects in steel has emerged as a prominent area of research. This paper introduces an innovative model, termed MPA-YOLO, which is based on YOLOv8 and aims to enhance the accuracy of steel surface defect detection. To improve the model's feature extraction capabilities, this study integrates and innovates upon large kernel depthwise convolution and coordinate attention mechanisms, resulting in the design of a multi-path convolution attention module (MPCA). Furthermore, MPCA is combined with C2f to create C2f-MPCA, which replaces parts of the backbone and neck network's C2f, thereby increasing the model's sensitivity to defect locations. Additionally, a partial self-attention module (PSA) is incorporated into the backbone network to capture long-range dependencies among features, thereby enhancing the representational capacity of the features. An auxiliary detection head is also introduced to gather multi-level and multi-scale feature information, which enables the model to effectively differentiate between target defects and background, thus improving its perceptual capabilities. The proposed model was assessed using the publicly available NEU-DET dataset, with experimental results indicating that the mAP of the MPA-YOLO model reached 81.5 %, reflecting a 3.4 % improvement over the baseline model. Concurrently, precision and recall rates increased by 3.0 % and 4.7 %, respectively. Furthermore, evaluations on the VOC2007 public dataset revealed a 2.5 % enhancement in mAP compared to the baseline model. These findings suggest that the MPA-YOLO model is effective in the detection of steel surface defects.