Partial Occlusion Handling using Quadrant-wise Sparsity for Face Image Recognition

In the past few years, Sparse representation technique for classification of face images has provided very good classification accuracy. The main problems associated with unconstrained face recognition are partial occlusion and expression changes. In case of partial face occlusion, a large portion of discriminative information gets lost, which deteriorates the classification accuracy. In this paper a novel quadrant-wise sparsity-based classifier is proposed which is based on implementing sparse representation in two different quadrants simultaneously. Proposed classifier utilizes the facial information available in upper and lower quadrants independently to perform classification of test image. Thus, it provides correct classification even if one quadrant is completely occluded. The efficiency of proposed method is showcased using detailed experiments carried out on standard AR database. The experimental results of different experiments carried out with different number of training images and training sets are compared with simple sparse representation method in terms of mean classification accuracy. The performance of proposed method is observed with a maximum improvement of 18.82 % in terms of classification accuracy.