Research On Face Posture Correction Algorithm Based On Multi-feature Fusion Network

In recent years,video surveillance has become the main technical means to ensure urban safety,and the recognition of facial images captured by surveillance equipment has also become an important part of facial image processing.At present,the recognition technology of a frontal face image has achieved excellent results and widely used,but in non-cooperative and non-sensing video surveillance scenarios,one or more profile face images with pose changes are often used for face recognition.A solution to this problem is to convert multiple profile face images into frontal face image,and then apply face recognition technology to recognition.In this application background,this paper aims at the reconstruction of the face image in two different application scenarios of multiple profile face images and profile face sequence images.The multiple profile face images have the characteristics of correlation and complementarity,and frontal face image reconstruction method based on deep fusion network is established,which provides the basis for subsequent face recognition.The main researches are as follows:(1)For frontal face reconstruction of multiple profile face images,a face posture correction method based on a multi-channel fusion generation adversarial network is proposed.The network consists of generator and discriminator.The generator takes multiple profile face images as input and uses multiple convolutional layers to sequentially extract the features of each image.In order to fully integrate the features of multiple profile face images,a convolution long-short-term memory network is introduced into a multi-channel fusion module,which uses multi-branch fusion to fuse the extracted multi-pose features to explore the relevant information and complementary features between multiple profile faces;the discriminator discriminates and analyzes the synthesized frontal face samples and uses the adversarial mechanism to adjust the network parameters.The network uses a multi-channel structure to fuse multi-pose features within and between layers to obtain the complementary features and global structure of the profile face image.The experimental results show that the proposed method can obtain the frontal face image with clear contour,which is beneficial to subsequent recognition applications.(2)For face reconstruction of profile face sequence images,a face posture correction method based on a recurrent feedback multi-fusion network is proposed.This method uses a two-dimensional convolution and three-dimensional convolution to sequentially obtain the spatial and complementary features of profile face sequence images,and enhances the fusion features by using multiple fusion methods.Before performing the three-dimensional convolution,the multi-pose features need to be recombined,so that the three-dimensional convolution can better capture the complementary features between profile face sequence images.In addition,the cyclic feedback mechanism is used to assist the current frontal face reconstruction process,which use the frontal face image reconstructed from the previous profile face sequence image as a priori information.The experimental results show that the frontal face image recovered by the proposed method not only has clearer face contour,but also has more detailed features.And the recognition rate of the synthesized face image can be higher.