Research On Pose Normalization Face Recognition Based On Generative Adversarial Network

In recent years,face recognition has aroused the research interest of researchers because of its advantages of nature,non-contact,imperceptibility and strong interaction.However,in complex and varied application scenarios,many factors can affect face recognition,such as occlusion,pose,expression and illumination.Among them,the change of head posture is one of the key factors affecting the recognition efficiency.At present,the methods proposed by researchers for pose face recognition can be roughly divided into three categories: pose robust feature extraction method,multi-view subspace learning method,and 2D or 3D face synthesis method.The premise of the first two methods is that the face pose training data is sufficient,which is often difficult to achieve in the actual scene.The third type of methods synthesizes fake faces for recognition tasks,which loses the fine texture of faces and lacks the robustness of face features under the change of pose.Therefore,these methods still fail to solve the problem of the degradation of recognition performance under the change of various pose.In order to improve the accuracy of pose face recognition,a pose normalization face recognition method based on generative adversarial network is proposed in this thesis.With pose change as the theme,experiments are carried out around the directions of face posture data expansion,landmarks detection and front face synthesis.The specific research contents are as follows:1.In this thesis,we propose pose face recognition method based on Pose-Guided Double Discrimination Generative Adversarial Network.Firstly,we design a pose-guided strategy to obtain pose templates through K-means algorithm and 3D Morphable Model(3DMM).Then,we propose a multi-image generation network,which can be utilized to extract the weighted features through pose templates and obtained the fusion feature representation.Besides,the self-attention module is designed to capture the important local features in detail under the pose variation.Especially,due to the assumption that the identity and the pose information play the different role in face synthesis,we design a dual discriminator network to guide the generator.As the results show,the proposed method can dramatically improve face frontalization appearances and enhance the precision of face pose recognition simultaneously.Exhaustive experiments on the Multi-PIE dataset and LFW dataset demonstrate that the effectiveness of this way in face recognition by synthesizing high quality positive face images.2.In this thesis,we propose a pose face recognition method based on Prior Symmetry Generative Adversarial Network.Firstly,Vgg-Face2 network is introduced to extract the identity features of the input image and eliminate the non-identity features,while the training of the generation model is supervised.Secondly,due to the prior symmetry of human face,the geometric symmetry extraction module is designed.On the one hand,it improves the local texture quality and global feature robustness of the synthesized face.on the other hand,it stabilizes the training of the network and speeds up the convergence speed of the model.Experiments are carried out on several common face databases,and the visualization results show that the method improves the quality of the synthetic image and achieves good recognition results.