Abstract:
Face recognition is an important area of research due to its requirement in our day-to-day life, be it surveillance or authentication. Current advancements in technology and computational power have shown promising results to solve this problem. Despite such furtherance, face recognition under uncontrolled environment still remains a challenging task and many state-of-the-art algorithms are unable to serve the purpose due to several challenges including varying illumination, pose, resolution, and occlusion. One primary reason for low performance is difference in training and testing data distribution. In this research, we propose an algorithm for face identification with varying pose and illumination. We propose an adaptive dictionary learning framework with Group Sparse Representation based Classifier to learn domain invariant dictionary representation of the given data. The algorithm adapts the representation learnt from the source domain with respect to the target domain in order to reduce the differences arising due to changes in the training and testing data distributions. Further, the data may contain noise and affect the dictionary atoms and group sparse coefficients, thereby hindering the discriminative power of the learnt dictionary. We propose to solve this problem using low rank minimization on dictionary atoms and group sparse coefficients. The effectiveness of the proposed algorithm is evaluated on the CMU MultiPIE and Extended YaleB face datasets for varying pose and illumination.