Face Recognition via Sparse Representation

Abstract

In this paper, we consider the problem of automatic face recognition form frontal view having different illumination and expression. The recognition problem can be presented as finding sparse representation of test image w.r.t training image. By computing sparse representation using l1 norm minimization we can generalize the algorithm for object recognition. Sparse provides insight in two important issues i.e. feature extraction and robustness to occlusion. In Feature extraction, the choice of feature is not critical as far as sparsity is properly exploited, only crucial requirement is that number of features should be sufficiently large. Error due to occlusion can be handled by using the fact that occluded pixels are sparse w.r.t standard pixel. Thus the sparse theory help us to predict how my occlusion the algorithm can handle. As the required number of feature in harnessing sparse is considerably large so we will extend the approach to application where only few no of feature or just one training image per atom is required for recognition. For this Extended sparse representation based classifier (ESRC) is used which apply an intra-class variant dictionary to represent variation between training and test image. Further to increase the robustness we considered robust sparse coding for face recognition which uses MLE (maximum likelihood estimation) solution of sparse coding problem assuming that the sparse coefficient follows laplacian or Gaussian distribution. Using MLE we can achieve better robustness to outliers (such as variation due to lighting, expression changes, disguise, and occlusion) than SRC. Here we are preforming sparse coding by Sparsity constrained robust regression problem for that an iteratively reweighted sparse coding algorithm is used to solve the Robust Sparse coding (RSC) model. This sparsity constraint on coding coefficient makes the computational cost of SRC too high. Assuming that coding coefficient and coding residuals are independent and identically distributed, we will robustly regress the signal with regularized coding coefficient using regularized robust coding regularized robust coding (RRC) Model. To solve this model an iteratively reweighted regularised robust coding (IR3C) algorithm is used. And this model shows that the sparse obtained using l1 or l2-norm almost give the same results.
In this paper we also verified dictionary learning method using fisher discriminant criteria. For sparse classification the dictionary used for coding plays an important role in it. Here we proposed a structured dictionary whose atom have correspondence to class labels and the reconstruction error will be used for pattern classification. As fisher discrimination criteria is applied so within class scatter will be minimized and between class scatter would be large. Thus the proposed dictionary learning method will used the discriminative information from the coding residuals as well as the sparse coefficients. All these method has been evaluated by using the AR face Database having different lighting or illumination, expression changes, occlusion, etc.