关键词: 口罩检测, 深度学习, YOLOv4, MobileNetv3, 深度可分离卷积

Abstract:

To overcome the drawbacks of the current YOLOv4 target detection algorithm, such as large network model, insufficient feature extraction, and susceptibility to light environmental impact, a lightweight YOLOv4-Lite network model with optimized feature extraction network and general convolution block is presented. Replacing the original trunk feature extraction network with MobileNetv3, which reduces the parameters of the network model and improves the detection accuracy. A deep detachable convolution block is proposed to replace the common convolution block in the original network, which further reduces the amount of parameters in the model. Combining label smoothing and learning rate cosine annealing decay algorithm, a new SiLU activation function is added to replace the ReLU activation function of MobileNetv3 shallow network, which optimizes the convergence effect of the model. Mosaic data enhancement method is optimized to improve the robustness of the model. In the face mask wearing task, compared with the original algorithm, the model sacrifices 1.68% of the mAP, but improves the detection efficiency（FPS） by about 180%.

Key words: mask detection, deep learning, YOLOv4, MobileNetv3, depthwise separable convolution