Deep Face Feature Extraction And Recognition

Face Recognition is received attention and in rapid development compared with other biological identification methods,which is mainly due to the universal,convenient and passive advantages of face recognition algorithm.However,face recognition is not applicable to unconstrained scene due to many reasons including human face is non-rigid object,diversity of human face poses,captured face image is vulnerable to environmental change and the limitation of computation capability and storage space.Therefore,overcoming the influence of pose variation and speeding up the algorithm is key requirements of face recognition algorithm.Face recognition with deep learning achieved encouraging progress,but several of them tackle the pose variation problem.Then,face identification with cosine similarity will take a lot of times if database images scale is large.So,taking deeper research on pose-robust features and faster algorithm is necessary.Finally,face presentation extracted by the proposed face recognition algorithm is robust to pose variation and take less time in face identification.The main research work and results are as follows:Conducting research about face recognition algorithm with convolutional neural network.Learn and summary the progress of face recognition algorithms in recent years,especially focus on the feathers,problem-solving methods and remaining problems.Learning basic procedures of face recognition algorithm and proposing pose related procedures including facial landmarks detection,face alignment,face pose estimation,face pose classification.Pose estimation and face pose classification are used to sense the pose of face in image.First,proposed deep feature transformation learning is able to transform any pose face representation into the corresponding frontal face presentation.Then,to learn the relationships between different poses and the corresponding network parameters,siamese network structure is used and corresponding loss function is proposed.Next,siamese network is converted into normal network through multi-task learning considering the high time complexity cost of siamese network.Last,pose loss function and pose center learning strategy is proposed to learn discriminative and pose-robust face presentation.Proposed deep hash algorithm for face recognition is able to lower the time complexity of face recognition,which combines feature extraction and binary hash two procedures into binary feature extraction.Through proposed hash layer and hash loss function in convolutional neural network,binary face presentations could be extracted.Considering the time complexity and performance of face recognition algorithm,cascade structure is chosen to combine deep feature transformation learning and deep face hash.Binary features are used to select gallery samples,which are sorted by similarity and top-K sorted gallery samples are selected to the next stage.Pose aware features are used to identify face with high precision in the next stage.