Research on Mask Wearing Detection Algorithm in Complex Scenes

Abstract

In recent years, various viral outbreaks have had a huge impact on the world, and wearing masks has become an important public health measure. Incorrect wearing of masks can increase the risk of virus transmission, so how to detect and monitor the correctness of mask wearing has become an important research direction. However, complex factors such as occlusion, dense crowds, and small-scale targets often exist in natural scenes, which have an impact on the detection performance of wearing masks. To address this issue, a mask wearing detection algorithm incorporating attention mechanism is proposed based on the RetinaNet algorithm. Specifically, it is to establish a branch in the residual learning unit of the original target detection network RetinaNet, and incorporate an attention mechanism with feature recalibration characteristics into this branch. The idea of this attention mechanism originates from the human visual system. In deep learning, it can be understood as obtaining the importance of each feature map channel through feature learning, and then assigning weights to each channel based on this importance, thereby achieving the effect of enhancing useful information and suppressing useless information. Due to severe facial occlusion in densely populated scenes, incorporating attention mechanisms into the object detection network can better learn local features of the target by enhancing important information weights, thereby achieving better detection performance. The experimental results show that compared with YOLOv3 algorithm, the detection accuracy of this algorithm on facial targets and facial mask wearing targets is improved by 0.11% and 3.16%, with an average detection speed improvement of 2 FPS.