A lightweight detector based on attention mechanism for aluminum strip surface defect detection

Many problems associated with the visual inspection of surface defects on aluminum strips remain to be solved, including the inapplicability of large-scale algorithm and computing equipment on site, and the balance between detection speed and accuracy. This paper proposes a novel and lightweight detection method based on attention mechanism, and focuses on the industrial application of aluminum strip defect inspection. On the basis of the YOLOv4 framework, the backbone network YOLO-DCSAM is constructed to utilize depthwise separable convolution and to design a parallel dual-channel attention module. It compresses the network scale and better enhances the effect of different channels on the feature map. At the same time, the neck network is redesigned and lightweighted for feature fusion, which can increase the receptive field and further simplify the network through SPPM-PANet module. Moreover, by optimization measure, such as the anchor box size of the cluster and improved loss function, the pertinence of model is strengthened to defect objects. The proposed method is trained and tested on the straightening aluminum strip surface data collected from the cold rolling workshop of Liuzhou Yinhai Aluminum Co., Ltd. Experiments show that the proposed method achieves a mAP of 96.28%, thereby outperforming the original YOLOv4 model. Moreover, as compared with YOLOv4, the model volume is reduced by 83.38% and the detection speed is increased by 3 times, thereby exhibiting the potential for real-time detection on the embedded systems.