Representation Learning And Regression Model For Robust Face Recognition

Face recognition is a hot topic in patter recognition and computer vision field.For decades,lots of progress have been made in face recognition techniques,which achieves satisfying performance when handling clean,clear and well-aligned frontal face images under the controlled environment.However,the existing techniques per-form poorly when dealing with low-quality face images,e.g.,images with variations of occlusion,low-resolution,misalignment and pose.This paper aims to study the robust face recognition methods to address the above mentioned variations,and the main contributions are summarized as follows:(1)A novel method named Learning Discriminative Singular Value Decomposi-tion Representation(LDSVDR)for face recognition is proposed.Recently,Singular Value Decomposition(SVD)based representation methods have attracted researchers'attentions for their power of alleviating the facial variations of occlusion,illumination and so on.The SVD representation reveals that the SVD basis set is important for the recognition purpose and the corresponding Singular Values(SVs)are regulated to form a more effective representation image.However,the existing SVD based repre-sentation methods all empirically make a rule to regulate the SVs,which is obviously not optimal in theory.To address this problem,this paper learns a common set of SVs by taking account of the information in the basis sets according to a discriminant criterion across the training images.Extensive experiments on popular face databases demonstrate the superiority of our proposed method over the previous empirical based SVD representations.(2)A Deep Recursive Residual Network(DRRN)for image super-resolution and face recognition is proposed.Existing face recognition methods suffer from the low-resolution face images.Therefore,effective face super-resolution method can signifi-cantly improve the performance when handling low-resolution face images.Recently.Convolutional Neural Network(CNN)based models have achieved great success in Single Image Super-Resolution(SISR).Owing to the strength of deep networks,these CNN models achieve impressive results at the cost of requiring enormous parameters.Compared to the compact models,large models demand more storage space and are less applicable to mobile systems.To address this issue,we build a deep vet concise network structure.Specifically,residual learning is adopted,both in global and local manners,to mitigate the difficulty of training very deep networks;recursive learning is used to control the model parameters while increasing the depth.Extensive bench-mark evaluation shows that DRRN significantly outperforms state of the art in SISR,while utilizing far fewer parameters.(3)A novel method named Structure Constraint Coding(SCC)for face recognition with image misalignment is proposed.Recently many sparse representation based face recognition methods with simultaneous misalignment were proposed and showed interesting results.These methods all use vector-based model and need to convert the 2 Dimensional(2D)image to the vector,which may lose some structural information.To address this problem,our proposed method skips this procedure and introduces a matrix regression model that uses nuclear norm to directly depict the structure error information in the original 2D image.Experimental results demostrate that compared with the sparse representation based methods,SCC is more robust for dealing with face recognition with image misalignment.(4)An Orthogonal Procrustes Regression(OPR)for face recognition with pose variations is proposed.Existing linear regression-based methods are sensitive to facial pose variations.To address this issue,we introduce the Orthogonal Procrustes Prob-lem(OPP)as a model to handle pose variations existed in 2D face images.OPP seeks an optimal linear transformation between two images with different poses so as to make the transformed image best fits the other one.To address the problem that the linear transformation is not suitable for handling highly non-linear pose variation,we further adopt a progressive strategy and propose the stacked OPR.As a practical framework,OPR can handle face alignment,pose correction,and face representation simultane-ously.Experimental results on popular face databases demonstrate the effectiveness of the proposed method.