Face recognition: Sparse Representation vs. Deep Learning

The pose, illumination and facial expression discrepancies between two face images are the key challenges in face recognition. The deep Convolutional Neural Networks (CNNs) and the fast Sparse Representation-based Classification (SRC) have achieved promising results in face recognition. However, CNNs require large databases and extremely expensive computations to overcome other algorithms. In this paper, we propose a novel SRC-based algorithm using test input image sets and training sub-databases, and compare its performance with CNNs. Histograms of Oriented Gradients (HOG) descriptors are used to define a new technique, named Training Image Modification (TIM), which provides image training sets with large variations of faces. The proposed algorithm divides the image training set into a number of sub-databases to address the dimensionality problem, and uses a test input image set to extract a signature from each sub-database using SRC. Each signature contains the same number of images as the test image set, although these may belong to different subjects. Considering all the sub-databases sequentially, the algorithm uses the signature of each sub-database to compute the number of images belonging to each subject. The signature that produces the Maximum Number of Images (MNI) of the same subject will have captured this subject for identification. YouTube Celebrity (YTC) and Multi-PIE databases are used in this work to evaluate the efficacy of the proposed method, which achieves high recognition rates. For relatively small databases, the proposed method is simple, scalable and stable, and it results in good face recognition rate under large face variations, as demonstrated by comparison with CNNs.