DICTIONARIES AND MANIFOLDS FOR FACE RECOGNITION ACROSS ILLUMINATION, AGING AND QUANTIZATION

Abstract

During the past many decades, many face recognition algorithms have been proposed. The face recognition problem under controlled environment has been well studied and almost solved. However, in unconstrained environments, the performance of face recognition methods could still be significantly affected by factors such as illumination, pose, resolution, occlusion, aging, etc. In this thesis, we look into the problem of face recognition across these variations and quantization. We present a face recognition algorithm based on simultaneous sparse approximations under varying illumination and pose with dictionaries learned for each class. A novel test image is projected onto the span of the atoms in each learned dictionary. The resulting residual vectors are then used for classification. An image relighting technique based on pose-robust albedo estimation is used to generate multiple frontal images of the same person with variable lighting. As a result, the proposed algorithm has the ability to recognize human faces with high accuracy even when only a single or a very few images per person are provided for training. The efficiency of the proposed method is demonstrated using publicly available databases and it is shown that this method is efficient and can perform significantly better than many competitive face recognition algorithms. The problem of recognizing facial images across aging remains an open problem. We look into this problem by studying the growth in the facial shapes. Building on recent advances in landmark extraction, and statistical techniques for landmark-based shape analysis, we show that using well-defined shape spaces and its associated geometry, one can obtain significant performance improvements in face verification. Toward this end, we propose to model the facial shapes as points on a Grassmann manifold. The face verification problem is then formulated as a classification problem on this manifold. We then propose a relative craniofacial growth model which is based on the science of craniofacial anthropometry and integrate it with the Grassmann manifold and the SVM classifier. Experiments show that the proposed method is able to mitigate the variations caused by the aging progress and thus effectively improve the performance of open-set face verification across aging. In applications such as document understanding, only binary face images may be available as inputs to a face recognition algorithm. We investigate the effects of quantization on several classical face recognition algorithms. We study the performances of PCA and multiple exemplar discriminant analysis (MEDA) algorithms with quantized images and with binary images modified by distance and Box-Cox transforms. We propose a dictionary-based method for reconstructing the grey scale facial images from the quantized facial images. Two dictionaries with low mutual coherence are learned for the grey scale and quantized training images respectively using a modified KSVD method. A linear transform function between the sparse vectors of quantized images and the sparse vectors of grey scale images is estimated using the training data. In the testing stage, a grey scale image is reconstructed from the quantized image using the transform matrix and normalized dictionaries. The identities of the reconstructed grey scale images are then determined using the dictionary-based face recognition (DFR) algorithm. Experimental results show that the reconstructed images are similar to the original grey-scale images and the performance of face recognition on the quantized images is comparable to the performance on grey scale images. The online social network and social media is growing rapidly. It is interesting to study the impact of social network on computer vision algorithms. We address the problem of automated face recognition on a social network using a loopy belief propagation framework. The proposed approach propagates the identities of faces in photos across social graphs. We characterize its performance in terms of structural properties of the given social network. We propose a distance metric defined using face recognition results for detecting hidden connections. The performance of the proposed method is analyzed on graph structure networks, scalability, different degrees of nodes, labeling errors correction and hidden connections discovery. The result demonstrates that the constraints imposed by the social network have the potential to improve the performance of face recognition methods. The result also shows it is possible to discover hidden connections in a social network based on face recognition.