Research And Application Based On Improved YOLOv4 Lightweight Network

As a popular direction in the field of computer vision,it is of great practical significance to solve target localization and recognition problems in computer vision by target detection algorithms.Among them,the target detection lightweight network has the advantages of streamlining,saving computational resources and easy deployment,but also has the problems of poor accuracy and poor detection of small targets.In this paper,we analyze the basic principle of YOLOv4 Tiny lightweight network,improve it for the above problems,and apply the improved network to the commodity detection task,and the main work of this paper is as follows.Firstly,this paper proposes the W-YOLOv4 Tiny network for the problem of low detection accuracy of YOLOv4 Tiny network,and improves the activation function of the backbone network as RRe LU to make the network dynamically adjust the parameters to adapt to different size data,so as to improve the generalization ability of the network.To address the problem that the deep semantic feature information is not fully utilized,a double branch structure is introduced into the original feature pyramid structure to enhance the feature fusion effect by fusing deep and shallow features,and the residual branch of the original feature layer is introduced to integrate features with the fused features to improve the feature fusion capability of the network.Secondly,to address the problem that YOLOv4 Tiny network has poor detection effect on small targets,it is improved to WL-YOLOv4 Tiny network based on W-YOLOv4 Tiny,and fused with CA attention mechanism to enhance key information extraction.The improved YOLO Head is a two-branch structure,which makes the classification and localization tasks simultaneously to improve the network robustness,and introduces a global average pooling layer in the classification branch to fuse the global information to improve the detection effect of small targets.Finally,to verify the effectiveness of the improved network in this paper,ablation,comparison and network visualization experiments based on the improved YOLOv4 Tiny network are conducted on the Pascal VOC dataset,respectively.The results show an overall improvement of 3.49% in the detection accuracy of the improved network on the Pascal VOC dataset to 80.40%.To verify the comprehensive performance of the improved network under the commodity detection task,the same experiments are conducted on the homemade commodity dataset.The experimental results show that the improved network improves the overall detection accuracy on the homemade commodity dataset by 8.73% to90.36% and the detection frame rate(FPS)is as high as 155.87.Compared with other advanced target detection algorithms,the improved network in this paper is most suitable for the commodity detection task.