Lightweight Mask Detection Algorithm Based on Improved YOLOv4

Abstract

In view of the fact that the current YOLOv4 target detection network structure is complex and difficult to be applied to low-computing hardware platforms such as mobile terminals, a lightweight mask detection algorithm based on improved YOLOv4 is proposed. Using GhostNet to replace the backbone feature extraction network of YOLOv4, the number of network parameters is greatly reduced and the detection speed is improved. The ordinary convolution of YOLOv4 is replaced with a depthwise separable convolution to further reduce the amount of model parameters. It is proposed to introduce the CBAM attention mechanism into the feature fusion network part to enhance the feature extraction ability of space and channel. The label smoothing and learning rate cosine annealing decay algorithms are used to optimize the convergence effect of the model, and Mosaic data is used to enhance the robustness of the model. Experiments on the constructed face mask dataset show that compared with the original YOLOv4, the proposed algorithm sacrifices 1.92% of mAP, the detection speed (FPS) is increased by about 71%, and the size of model is reduced by 201. 86M. Compared with other mainstream target detection algorithms, this algorithm also has better detection effect and can meet the accuracy and real-time requirements of mask detection tasks.