Robust face recognition via hierarchical collaborative representation

Collaborative representation-based classification (CRC) is currently attracting the attention of researchers because it is more effective than conventional representation-based classifiers in recognition tasks. CRC has shown high face recognition accuracy; however, its accuracy is degraded significantly if the number of training faces in each class is small. This is because the accuracy of CRC is only dependent on the results of minimizing the Euclidean distance between a testing face and its approximator in the collaborative subspace of training faces. In this research, we proved that the accuracy of CRC can be improved substantially by minimizing not only the Euclidean distance between a testing face and its approximator but also the Euclidean distances from the approximator to training faces in each class. Consequently, we presented a hierarchical collaborative representation-based classification (HCRC) in which a two-stage classifier is applied for training faces, and the recognition accuracy of the second-stage classifier is significantly improved in comparison to that of the first-stage classifier. Moreover, the recognition rate of our classifier can be considerably increased by using models of discriminative feature extraction. Since noise and illumination are the main factors that cause CRC to be less accurate, we propose combining HCRC with a wide model of local ternary patterns (LTP). This combination enhances the efficiency of face recognition under different illumination and noisy conditions. For dealing with face recognition under variations in pose, expression and illumination, we present a deep convolutional neural network (DCNN) model of discriminative feature learning, which transforms face images into a common set of distinct features. The combination of HCRC with this deep model achieves high recognition rates on challenging face databases. Furthermore both models are optimized to reduce computational costs so that they can be successfully applied for real-world applications of face recognition that are required to run reliably in real time. In addition, we also prove that combining state-of-the-art DCNN models with HCRC results in an significant improvement in face recognition performance. We demonstrate several experiments with challenging face recognition datasets. Our results show that the hierarchical collaborative representation-based classifier with the models significantly outperforms state-of-the-art methods.