Research On Key Technologies Of Face Recognition For Video Surveillance

The rapid development of artificial intelligence has been widely used in the field of computer vision,making intelligent image processing algorithm greatly developed.Face recognition,as a key technology in computer vision and image processing,is one of the hot research topics in academia in recent years.At present,face verification and face recognition in restricted scenes has been basically mature,and has been widely used in public security,finance and social services and other important fields.However,in unrestricted scenes represented by video surveillance system,faces are easily affected by posture,expression,illumination or occlusion,resulting in facial feature defects,which greatly reduces the accuracy and real-time performance of face recognition in unrestricted scenes.In light of the shortcomings of the existing key technology and facing the challenges of in-depth analysis and research,the research mainly studies the cross-domain face recognition,cross-pose face recognition,lightweight and training of the recognition model in order to improve the unrestricted scene oriented face recognition accuracy and real-time performance.This thesis proposes a robust and real-time solution for face recognition.The main research contents and innovations of the thesis are as follows:1.There is a domain gap between the training samples in the public data set and the practical application environment,so the face recognition model trained on the public data set cannot be effectively generalized to the real application environment.At present,the identity information of face image cannot be retained effectively when implementing image style transfer.To solve this problem,a face image domain transfer technology based on domain adaptation is proposed.The technique uses unpaired adversarial training to learn the feature mapping between the source domain and the target domain,and transfers samples from large open data sets to the target domain.The transferred images have the same statistical features as the test environment.Based on domain adaptation,identity objective function is added to improve the retention degree of identity features in the process of domain transfer.On this basis,the face identity feature retention in domain transfer is tested on public face databases,and the performance of cross-domain face recognition is verified by this technique.2.In order to solve the problem that additional recognition networks need to be introduced in face recognition based on frontal face generation,a multi-pose face recognition technology based on Cross-Pose Generative Adversarial Network(CP-GAN)is proposed.The method corrects multi-pose faces by learning the mapping between profile faces and front faces in image space.For single image and video frames,multi-pose face correction algorithms based on single image and multiple image are studied respectively.After realizing multi-pose face adjustment,face image recognition results can be obtained directly through weighted-sharing convolutional neural network,which is an end-to-end pose invariant face recognition method.Then,the validity of the proposed method is verified by face verification on CFP dataset and face recognition experiment on CASIA 3D Face,IJB-A and Multi PIE datasets.Experimental results show that compared with other PIFR methods,the proposed method has better face verification matching rate and recognition accuracy,and better robustness and generalization ability in unconstrained scenarios.3.Aiming at the difficulty of deploying deep learning model caused by limited processing capacity and small memory of terminal video acquisition equipment,a deep neural network lightweight algorithm based on knowledge distillation and adversarial learning is proposed to transfer knowledge with classification probability and middle-layer feature.The principle of the algorithm is to improve the traditional knowledge distillation loss and add the indicator function,so that the student network only learns the classification probability correctly identified by the teacher network.The discriminator of adversarial learning strategy is introduced to identify the difference between student network and teacher network in middle layer.In the latter part of the training,the teacher network and the student network learn from each other and update each other alternately,so that the student network can explore its own optimal solution space,and further improve the generalization ability of the student network,so that it can be applied to different machine vision tasks.The experimental results show that the small size student network obtained by knowledge distillation has less parameter,and the recognition accuracy decreases only about 1.5%.4.Aiming to solve the problem of high computational resource and bandwidth consumption of training deep face recognition model,a distributed training based deep face recognition method is proposed.In this method,limited local data and model upload are adopted to realize global model training in cloud with less bandwidth consumption.Then the global model guides the update of the local model to improve the performance of the local model without occupying local computing resources.Experimental results show that compared with the local training method,distributed training has the advantages of less bandwidth consumption and saving local resources.