Robust Face Recognition Using Sparse and Dense Hybrid Representation with Local Correlation

Abstract

Face Recognition is one of the biometrics that can be used to uniquely identify an individual based on the matching performed against known faces. The real world face recognition is very challenging since the face images acquired may vary with illumination, expression and pose. No existing system can claim that they have handled all these issues well. This work particularly focus on addressing the problems of face images taken in challenging environments. A more efficient Face Recognition system based on a combination of Sparse and Dense representation (SDR) along with Local Correlation is proposed. While considering the efficient methods for classification, Sparse Representation (SR) is the best one. Here a Supervised Low Rank (SLR) decomposition of dictionary is used to implement the SDR framework in the initial step. Then we apply Local Correlation to the cases where SDR-SLR method fails to distinguish competing classes properly. Usually due to changes in illumination and pose, variations can be seen to occur in different face parts. Correlation is calculated between the query image and the images of top matches that are obtained from the SDR-SLR method. Since we compute local correlation of relevant points only within a small dictionary, computation time of the proposed method is very less. Challenging benchmark datasets such as AR, Extended Yale and ORL databases are used for testing the proposed method. Experimental analysis shows that performance of the proposed method is better than the state-of-art face recognition approaches and the performance gains are very high.