A study of the lightweight algorithm for recognizing masks incorporating attention mechanism and G-GhostNet

Abstract

As COVID-19 and humanity's protracted battle makes wearing masks a norm, intelligent recognition of mask-wearing and wearing criteria is investigated to lessen the labor of epidemic control and detection personnel. To achieve the purpose of not having to manually identify mask use in a busy, open setting, to be able to provide a real-time warning of the whole mask non-wearing phenomenon in mobile devices, and to address the issue of sluggish efficiency and poor precision of traditional target detection and tracking means in complex scenes, this work presents a lightweight mask identification method that combines G-GhostNet with an attention mechanism. Firstly, YoLov5's BackBone network is replaced with a more lightweight G-GhostNet with fewer convolution kernels to reduce the parameters and computation more significantly; secondly, an attention mechanism is introduced to G-GhostNets feature mixing process to increase discovered features weight values.Then, the dichotomous unmasked and masked datasets are prepared to enhance the model accuracy and speed by the loss function. The experiments compare the effects of Squeeze-and-Excitation Networks (SE), Convolutional Block Attention Module(CBAM), Efficient Channel Attention (ECA), and Coordinate Attention (CA), the four attention processes on the enhanced model with the model without the added attention mechanisms. The results show that adding the attention mechanisms all make the model recognize masks better, among which adding CA has the best recognition effect with a 2% improvement in accuracy but 1.3 times increase in Latency time; in contrast, adding SE, ECA balanced speed and accuracy.