Mask detection algorithm based on the improved YOLOv4 - tiny

Abstract

Based on YOLOv4-tiny, A lightweight mask detection algorithm is presented. By replacing the CBL module in the backbone feature extraction network (CSPdarknet-tiny) and Yolo Head with Ghost module that reduces the parameters of the network model. By the combination of Ghost module, CBAM attention, SMU activation function, and BN layer, a lightweight attention mechanism residual module (GCS_Block) is designed, which is embedded into the backbone feature extraction network, improving the model extract mask feature level. The Kmeans++ method is used to perform anchor box clustering on the dataset in this thesis. The experimental results show that compared with YOLOv4-tiny, the MAP has increased from 74.02% to 86.77%, the parameter has decreased from 6,056,606 to 1,657,828. The memory size of the model is 5.6MB.