Sparse Regression Based Face Recognition

Despite the large volume of research in the literature, face recognition remains a hard problem to solve. The challenge is due to many factors that may affect the performance of any recognition system such as size of training data, noisy images, accuracy-speed trade-off, variations in illumination, expressions, or pose, etc. Although there are a lot of efforts that have been proposed for special conditions, no method exists to work under unconstrained conditions with satisfactory performance. In computer vision, human face classification is a very popular and important topic. This popularity comes from the wide-spread applications of face recognition such as entertainment, security, and control. Most, if not all, of these applications require the recognition systems to have low computational complexity and high recognition accuracy. In this thesis, three methods are proposed for feature extraction and face classification. These methods are based on `2-norm regularized regression. The main idea of these methods is to train a dictionary capable of transforming a face image into a form that can be used to classify it into its correct class. Image representation in the transformation domain is a sparse vector with small number of nonzero coefficients. The goal is to come up with a transformation matrix such that the sparsity pattern depends on the class of the transformed image. This is accomplished by regression in addition to regularization terms and constraints. The three proposed ideas differ in how to attain this goal. The first proposed method uses the idea of predefined sparse matrix to specify the sparsity pattern of image transformation. The second method constrains the transformation such that the inner product of the image transformation is minimized if the images are of different classes, and maximized if the images are of the same class. The last method transforms the images such that the nonzero coefficients do not overlap for different class, and the transformation of class images become close to the transformation of its mean vector. Several simulation experiments are implemented and executed to test the performance and competence of the proposed methods. The simulations are performed with different benchmark face databases. The results prove that our proposed methods are distinguished over state-of-the-art methods by their accuracy, computational cost and robustness to image occlusion and corruption