Research On Mask Face Detection And Recognition Algorithm Based On Deep Learning

As novel coronavirus has swept the world,nucleic acid detection has increasingly become the most direct means of normalization of epidemic prevention and control.Motivated by big data,artificial intelligence,and other technologies,with the goal of more rapidly and accurately understanding people’s health,for current epidemic prevention policy,it is important to detect whether people are wearing masks and recognize faces under masks.However,because the face target with mask is small,it is easy to be affected by natural conditions in public scenes,and the mask occlusion makes the face features in the picture suffer huge losses,which is difficult for face recognition.In view of the above problems,we design a mask-wearing detection algorithm based on depth learning and a face-mask recognition algorithm respectively in this thesis.The major research contents and innovations of this article include:(1)The AIZOO mask wearing detection data set was preprocessed,and the Retina Net algorithm was trained and tested for the accuracy and speed,and compared to the other algorithms for the experiments and analyses.The selected Retina Net algorithm is proven to perform well on all indicators.Finally,the problem that the AIZOO dataset has few face masks in public scenes is analyzed.In order to further observe and analyze the image visualization effect of the model,200 images of face masks in public scenes are obtained through network downloading for detection,and to further test the model’s ability to generalize.(2)Aiming at the problems of Retina Net algorithm in mask wearing detection of small and medium-sized targets,low detection accuracy of single category in dense scenes and large impact on public scene area,RCCAF-Retina Net mask wearing detection algorithm is proposed.Res Net50,a feature extraction network,focuses more on interesting features by embedding CBAM attention mechanism to reduce the interference of features around the target on the target,the CA attention mechanism module is appended to the shallow feature layer of the FPN feature pyramid network,in order for the network to focus more on learning the details of the mask features and to strengthen the receptive field of the lower layer.Experiments show that the improved RCCAF-Retina Net algorithm does indeed improve the detection accuracy,and resolves the issues of false detection and missed detection to some degree.(3)In the face mask recognition task,the mask occlusion makes the face features in the picture suffer a great loss,adds useless mask features,and the recognition accuracy cannot meet real world requirements.In this thesis,we propose an RSGe M-Face Net face-mask recognition algorithm.Res Ne Xt50 is used as the base network,and the SENet channel attention module is embedded to modify the area weights,so that the face area not covered by the mask has a larger weight,while at the same time reducing the weight of the area covered by the mask and making more complete use of the area that is not covered by the mask.In addition,the Ge M generalized average pooling layer is used to obtain more mask-occluded facial features,and the feature dimension reduction is carried out to reduce the negative impact of the masks on the network’s recognition accuracy.Experimental results show that the enhanced RSGe M-Face Net algorithm has some improvement in the recognition function of the face masks as compared to the original algorithm.