Face Reconstruction And Recognition For Video

Video surveillance is currently one of the main security methods.Using video clips and pictures to perform accurate face recognition has an important role and significance in security,criminal investigation and other fields.In this application scenario,because the face image is usually obtained without cooperation,there are often problems such as attitude deviation,blur and occlusion,which leads to the failure of the existing face recognition algorithm,and the recognition accuracy rate seriously declines.application.Therefore,studying the face recognition method of video surveillance can not only serve as an important supplement to the existing face recognition scheme,but also expand the application range of face recognition,which has important research significance.This paper is oriented to non-cooperating scenes of video surveillance,and studies the face reconstruction and recognition based on a single picture,and the fusion and recognition of facial features based on multiple pictures.The main contents are as follows:(1)Aiming at the problem of the lack of high-frequency information in low-resolution and multi-pose face images,and the poor robustness of the reconstructed face,a ternary confrontation reconstruction recognition network is proposed.The network includes two parts: face reconstruction and face recognition.First,a codec network is used to reconstruct a low-resolution side face image into a high-resolution front face image,and then a face recognition network is used to perform the reconstruction of the reconstructed front face.Face recognition.In the network training,a shortest distance triplet loss function is designed,and the loss function is combined with the confrontation mechanism,and a ternary confrontation training method is proposed.The shortest distance triplet loss is applied to the reconstruction In the recognition network,the network is more similar to the features extracted by the same person,but not similar to the features extracted by other people,so that the facial features generated by the reconstruction are closer to the unique features of the real frontal face,providing robustness for subsequent recognition Characteristics.The experimental results show that the algorithm proposed in this paper can still reconstruct a clear and true frontal face image under the condition of low resolution and large bias,and obtain a higher recognition rate.(2)In order to extract key features that are easy to identify from face sequence images with redundancy,correlation and complementarity under video surveillance,a video face recognition method based on residual recurrent network is proposed.This method fully extracts and integrates the features of the time dimension and the space dimension of the video frame.After each frame of the image generates the prediction matrix sequence,the sequence is fused to produce the video face recognition result.In the feature fusion of sequence images,the residual loop mechanism is introduced into the video frame sequence fusion to avoid the disappearance of the contextual gradient of the video face during the feature fusion process,and effectively ensure the sequence length of the feature fusion.The experimental results show that compared with the feature fusion of 2D convolution,3D convolution and recurrent neural network,the video face recognition method of residual recurrent neural network can extract richer context information,and can better perform the preceding and following frames.The mutual compensation between them has better robustness and stronger recognition ability for long-sequence face videos,and is better than other video face recognition algorithms in time complexity,and has better real-time performance.