Research On Face Detection And Optimization Based On Surveillance Video

As face recognition technology matures,many products and businesses related to face recognition have been launched.In the video surveillance industry,there are many businesses that require face recognition.However,in an unconstrained monitoring environment,the face recognition system is easily affected by illumination,angle,occlusion and expression,etc.resulting in low-quality images obtained by face detection,such as non-face images and occluded face images.Especially under the current severe new crown epidemic prevention and control,mask occlusion has brought huge challenges to face recognition.Relevant studies have shown that these low-quality face images are an important factor affecting the performance of face recognition systems.In order to improve the accuracy of face recognition,it is usually necessary to evaluate the quality of multiple captured images of the same target,remove inferior images,and use high-quality images for subsequent face recognition tasks.Therefore,It is of great practical significance to study face detection and optimization in video surveillance scenes.The main work of this dissertation is as follows:Firstly,To solve the problem of high false detection rate in face detection in video surveillance scenarios.This thesis proposes a CNN-based face determination model.The model scores the confidence of the input image,and determines the face image through the threshold,so as to distinguish the face from the non-face.The experimental results show that on the self-built surveillance video dataset,compared with the Alex Net,VGG-16,and VGG-19 models,the model in this thesis has the highest accuracy and the least model parameters.Compared with the three networks of Google Net,Res Net,and Mobile Net,the accuracy rate is not much reduced,but the parameter scale is greatly reduced by 84.8%,66.4%,and 70.5%,respectively.On the self-built surveillance video data set,the two methods of face recognition after face detection and face recognition after face judgment are compared.The test results show that the latter is accurate in face recognition.The rate has increased by 14.8%,and the time-consuming has been reduced by about one-third.Secondly,To solve the problem of face occlusion detection in video surveillance scene,this thesis proposes a face occlusion detection method.The main idea of this method is to refer to the ratio of the number of occluded key points in a certain area of the face to the total number of key points in the area,calculate the degree of occlusion of the four areas of the face’s eyes,nose and mouth,and then obtain a total face by weighting The occlusion score is compared with the threshold to judge the occluded face.The experimental results show that on the self-built surveillance video dataset,compared with Faceness-Net,FAN,DSFD,C-GAN and Pyramid Box methods,the method in this thesis has the highest detection rate of face occlusion.On the basis of the MTCNN network,the comparison before and after adding the method in this thesis,the test results show that on the self-built surveillance video data set,the MTCNN network after adding the method in this thesis can improve the detection accuracy of occluded faces by 4.1%.Finally,this thesis develops a face image optimization system based on surveillance video.It can judge the face images and large-area occluded face images in the surveillance video,and comprehensively evaluate the face images based on the four factors of face image clarity,face occlusion degree,human eye status and face status and finally filter out the best results.