Novel Efficient Steel Surface Defect Detection Model Based on ConvNeXt v2 and Squeeze Aggregated Excitation Attention

With the rapid advancement of artificial intelligence and intelligent manufacturing technologies in recent years, automated inspection technologies have ushered in new opportunities for industrial production. Against this backdrop, this paper proposes an improved single-stage detection model, ESS-YOLO, which enhances the detection capability of steel defects in complex industrial environments, particularly for large and small target defects that are confounded by complex backgrounds and ambiguous defect locations. Firstly, by integrating the concepts of CSP architecture with ConvNeXt v2, a module is constructed that enhances the capability of image feature extraction. Secondly, SaE channel attention is employed to bolster the feature representation in the detection head, providing robust support for classification. In addition, to mitigate the issue of excessive resource requirements for the detection head, DSC is incorporated, which effectively reduces the complexity of the model. Finally, The Shape-IoU loss function enhances the model’s ability to learn the geometric properties of bounding boxes, taking into account both the geometric information of the ground truth boxes and the predicted boxes, thereby improving the accuracy of regression. Experimental results on the NEU-DET dataset demonstrate that, compared to the original model YOLO v8, the proposed ESS-YOLO model achieves a 2.1% increase in detection accuracy and a 19.5% reduction in computational load. Compared to common detection algorithms such as Faster-R-CNN, the accuracy is improved by 7.1%, indicating the excellent detection performance of our model. These results suggest that the ESS-YOLO model has significant application potential in the field of steel defect detection.