Abstract: In view of the large difference of different sizes of aluminum surface defects leading to poor detection effect, an aluminum surface defect detection algorithm based on improved YOLOv5 is proposed. Firstly, CA attention mechanism module is embedded in the network to make the network better suppress the interference of invalid samples in the image and focus more on useful information. Secondly, a small target detection layer is added to the original detection layer to acquire and transmit more abundant and discriminant small target features, so as to solve the problem of low detection accuracy of small target defects and improve the overall detection accuracy. Finally, the SIoU loss function is introduced and the vector angle between boundary box regression is introduced to redefine the loss function, which effectively reduces the total freedom of loss and improves the reasoning precision. The improved algorithm is applied to Tianchi aluminum data set for verification. The experimental results show that the model can effectively identify different types of defects on aluminum profiles, which is 11.4% higher than the original YOLOv5 algorithm mAP, and the detection speed is up to 66.4 frame/s. The proposed method can meet the requirements of on-site defect detection in aluminum profile factories.

Key words: defect detection, YOLOv5, attention mechanism, SIoU, multiscale fusion