Surface Defect Detection Method of Hot Rolled Strip Steel Based on Fusion Attention and Multi-scale Features

Aiming at the problems of small defect area, diverse shapes, blurred boundaries and complex backgrounds of surface defects of hot-rolled strip steel, a surface defect detection model for hot rolled strip steel, named SFSP-YOLOv7, is proposed. First, by improving the k-means++ clustering algorithm, the prior frame dimension was adjusted, the Euclidean distance metric was replaced by intersection over union (IoU) distance, and genetic algorithm (GA) was introduced to obtain a more representative anchor box size, improving the regression speed of the model and the precision of small area defect detection. Second, for the defects with fuzzy boundary and complex backgrounds, an object detection boundary frame loss function, named FocalSIoU, was proposed to reduce the learning of unnecessary features in the model, accelerate the detection speed, and improve the regression effect of the prediction frame. Finally, a multi-scale feature fusion module (MFFM) was designed to enhance the model's feature extraction capabilities through multi-scale information fusion, improve the detection precision of small targets, and improve the misdetection rate of model detection. The space to depth (SPD) convolution module was introduced into the model Head structure to improve the model to avoid the loss of fine-grained information and reduce the target miss detection rate. Through the verification of NEU-DET data set, the mean average precision (mAP) value of the SFSP-YOLOv7 model is 78.3%, which is 5.0% higher than the original YOLOv7 model, indicating the effectiveness of the detection method proposed.