Research On Post Robust Face Recognition Algorithms

Face recognition is one of the major topic in the fields of artificial intelligence,pattern recognition and computer vision.Face recognition is widely used in many fields,such as information security,electronic commerce and security access control.However,the large poses of faces have become one of the main obstacles that affecting the results of face recognition.In this thesis,we focus on the plane rectification method for large poses of human face and the robust feature representations of faces through the deep neural network model.The main innovations and work of this thesis can be summarized as follows:1)Pose invariant face recognition using maximum Gabor similarity based on the weighted Lucas-Kanade(LK)algorithm is proposed.Firstly,the warp parameter is obtained by using the weighted LK algorithm between the probe image patches and the gallery image patches,the criteria that the globally optimal warp parameter is obtained by using maximum Gabor similarity is proposed.Furthermore,the average Gabor similarity acquired from the globally optimal warp parameter is used as the face recognition weight,so that it can improve recognition rates and enhance the robustness of the pose invariant face recognition.Finally,the FERET database is used to evaluate the optimal parameters and rectify the profile face.At the same time,the recognition rate experiment is also carried out,which proves the effectiveness of the proposed method.2)Profile-to-frontal face transformation in the deep feature space based on multi-task algorithm is proposed.The Deep Residual Equivariant Mapping compensates the deep feature of profile face so as to realize profile-to-frontal face transformation in the deep feature space.A new task is added to the original Deep Residual Equivariant Mapping to reconstruct the input features of the profile face.The new task with the original residual transformation constitutes a multi-task.Through the role of multi-task,the generated face depth space features can be evaluated,and the recognition ability of the network can be improved,the whole network can be more robust.The experimental results show that the performance of recognition is improved by adding multi-task compared with the original module.