Improved Yolov5 Algorithm for Surface Defect Detection of Strip Steel

Aiming at the problems of low detection accuracy and slow detection speed in the traditional method of detecting strip steel surface defects, this paper proposes an improved yolov5 algorithm for detecting strip steel surface defects. Firstly, the data set of strip surface defects was constructed, and the K-means algorithm was used to cluster the defect samples, and the prior box parameters of different sizes were obtained. Secondly, the attention-yolov5 algorithm is proposed, which draws on the item-based Attention mechanism, adds channel Attention and spatial Attention mechanism to the feature extraction network, and uses the filtered weighted feature vector to replace the original feature vector for residual fusion. Finally, In order to improve the ability of defect feature extraction, the convolution layer is added after the main feature is extracted from different feature layers of the network output and after the pooling structure of spatial pyramid. The experimental results show that the mAP value of the improved yolov5 algorithm on the test set is as high as 87.3%, which is 5% higher than the original yolov5 algorithm. The average detection time of a single image is 0.0219s, which is basically the same as the original algorithm, and the detection performance is also better than the Faster RCNN and yolov3.