Research On Dynamic Face Recognition Based On Deep Learning

With the significant improvement of people's living standard and the increasingly strong demand for a better and safe working and living environment,video dynamic face recognition technology has been widely studied and highly concerned by all sectors of society.Because video data contains images that are unfavorable to face recognition,such as those caused by illumination change,pose change,occlusion and image blur,etc.,the results of dynamic face recognition are still in the primary stage,especially the poor robustness of recognition,which cannot meet the requirements of practical application.Compared with static face recognition,dynamic face recognition has higher theoretical research significance and practical application value.Fortunately,with the coming era of artificial intelligence,advanced deep learning technology provides a new solution to the dynamic face recognition problem in video.This paper proposes a dynamic face recognition method under video monitoring environment to solve the problem of poor robustness of dynamic face recognition.The main work and contributions are as follows:(1)This method is inspired by recent in-depth learning techniques and has designed a Recurrent Network of simultaneous Regression and Classification(srcr-net).Srcr-net combines the advantages of mobile learning,multi-task learning,augmented learning,and cyclic neural networks.The transfer learning strategy can reduce the training convergence time of the model.Multi-task learning can track face target and recognize face category at the same time,which is helpful to improve the effect of dynamic face recognition.Enhanced learning emphasizes the learning weight of facial focus area and improves the accuracy of recognition.The cyclic neural network has certain memory information ability,which is suitable for video image information and can improve the robustness of dynamic face recognition.Firstly,video face image features of each frame were extracted with pre-trained model,and then simultaneously tracked and identified.Then,reinforcement learning was used to further improve the recognition effect,and finally,a segment of video was identified by using circular neural network.The proposed network model adopts the strategy of online training and offline testing,which can improve the speed of face recognition and meet the realtime requirements of the dynamic face recognition system.(2)Under the conditions of video monitoring,easy to generate images for face detection,in this paper,the traditional Adaboost method of face detection is improved,before detection of human face skin color model to create and facial skin brightness generalization model creation,is used to locate face region in advance,this method can effectively alleviate the complex background image and strong illumination change to the negative influence of face recognition.(3)Design experiments to verify the srcr-net model and its deformation model,and analyze the influence of migration learning,multi-task learning,reinforcement learning and cyclic neural network on improving dynamic face recognition effect;At the same time,the proposed method is compared with the current advanced method.According to the experimental results,the recognition accuracy of the proposed method is close to the best performance and the robustness is greatly improved.The effects of the distribution and size of the experimental data,different loss functions and different layers of LSTM on the model identification results are also analyzed.(4)The deep learning knowledge and face recognition technology are summarized.This paper briefly introduces the relevant concepts of deep learning and explains in detail the principle of characteristic learning ability of deep learning,as well as the relevant theoretical knowledge of convolutional neural network.The advantages and development trend of face recognition based on deep learning are summarized and analyzed.In order to improve the public security ability and effectively protect the safety of people's lives and property,this paper designs a dynamic face recognition method under video monitoring with the help of the current hot wave of deep learning development.This method solves the technical difficulties of dynamic face recognition to a certain extent,especially effectively improves the robustness and real-time performance of dynamic face recognition.