Study On Pose-Robust Face Recognition

In recent years,deep learning methods are widely used in many research fields in computer vision,and it achieves great success in face recognition research.Now,the performance of face recognition methods in restricted environment is close to saturation.At the same time,the increasing of the quality and member of urban surveillance cameras makes the easy to acquire face image data in the wild.For this reason,the market application of face recognition technology is greatly expanded,such as the face recognition application system in unrestricted environment under security and monitoring scenario.And in this context,researchers began to select face recognition in unrestricted environment as the next study.Among this,the study on pose-robust face recognition under real scene is a hot and difficult research topic.Due to the rigid deformation of face image caused by the pose-changing,the traditional representation feature based face recognition methods are difficult to obtain the pose-robust feature.And the face self occlusion,illumination and facial features changings also increase the complexity of face image,which brings great difficulties to face recognition mission.In this thesis,we focus on the multi-pose face recognition problem from two aspects: robust feature extraction and frontal face image synthesis.And the main research work and innovation are concluded as below.First,in the research work of robust feature extraction,a multi-pooling feature fusion method is proposed by this work.This method extends the robustness of the features,which extracted by fusing different pooling features from the pooling layers in convolutional neural network.Meanwhile,in this research we establish the concepts of local feature and global feature.And the fusion feature obtained by local and global can effectively improve the recognition performance of deep learning methods with complex pose changing.Second,in the research work of frontal face image synthesis,we propose the Geometry Structure Preserving based Generative Adversarial Network,GSP-GAN,for multi-pose face frontalization and recognition.More importantly,for discriminator of our model,we use the self-attention block to preserve the geometry structure of a face.The discriminator consists of a series of parallel sub-discriminators that carry the global and local attention information.The proposed GSP-GAN can generate high-quality frontal images under arbitrary pose,which effectively solve the interference problem when face image under large pose changing.And it gets satisfactory recognition performance.Finally,the effectiveness of the two methods proposed in this research is verified through the experiments on the Multi-PIE face database.