Research On Face Recognition In Surveillance Scene Based On Deep Learning

Face recognition is one of the hot topics in pattern recognition and computer vision,which is widely used in security,transportation,finance and other fields.With the development of deep learning and big data technology,the performance of face recognition has been improved and surpassed the human level to a certain extent.However,in the surveillance scene,the performance of face recognition is not high because of the light,occlusion,posture,age and other reasons.Therefore,this paper focuses on the in-depth study of face recognition algorithm in surveillance scene,and proposes a face recognition algorithm in surveillance scene based on deep learning.The main tasks are as follows :(1)In view of the problem of intra class difference caused by age and the difficulty of collecting cross age face data in the surveillance scene,a face recognition training sample enhancement based on the generation adversarial network is proposed.The youth and old age data corresponding to the middle-aged face data in CASIAWebFace data set are generated by the generation adversarial network to realize the sample enhancement.In addition,in the training process,the generated data and the real CASIA-WebFace middle-aged data are classified to ensure that the generated face is similar to the person.The algorithm effectively solves the problem of difficult to collect face data across age groups and expands the limited training samples.(2)There is no intersection between the training samples and the test samples of face recognition in the surveillance scene,which belongs to the open set face recognition.Face comparison is needed.From the perspective of measurement learning,an improved additive cosine margin softmax loss function is proposed,which improves the additive margin softmax loss function.By subtracting a value from the cosine value of the angle between the feature and the target weight,the feature and the untargeted weight are improved The cosine value of the heavy angle plus a value,which is the number between 0 and 1.The best value is selected through experiments to reduce the intra class distance and enlarge the inter class distance,so that the trained model can extract more distinguishing features.(3)In order to solve the problem of face pose and age in the surveillance scene at the same time,a face recognition method based on multi task convolution neural network is proposed.The generated data and CASIA-WebFace data are input into the improved resnet-20 network to perform the task of face identity classification,face pose classification and face age classification.Among them,the task of face identity classification is the main task.The task of face pose classification and age group classification are secondary tasks.Improve the accuracy of multi pose cross age face recognitionTest on face verification,multi pose,cross age face test sets and self collected surveillance scene test sets.The experimental results show that the algorithm proposed in this paper improves the accuracy of multi pose and cross age face recognition in a certain extent,and improves the performance of face recognition in the surveillance scene.