Face recognition is an important branch in the field of biometric recognition. Compared with other biometric recognition, such as fingerprint and iris recognition, face recognition has two important advantages: non-compulsory recognition and large scale recognition. This means that a person can be identified by face recognition in the crowd, even if he or she is not cooperative. These advantages of face recognition make face recognition technology widely used in video surveillance, access control, law enforcement, etc. Collecting of face samples is one of the main difficulties for face recognition. In most of the real-world applications such as law enhancement, e-passport, and ID card identification, it is customary to collect a single sample per person (SSPP), i.e., the profile face image, which leads to SSPP face recognition problem. Unfortunately, in such SSPP scenario, many presented face recognition methods suffer serious performance drop or fail to work due to their inability to learn the discriminative information of a person from a single sample. Moreover, in video surveillance, as the installed cameras are far away from the pedestrians, face images captured by these surveillance cameras tend to be with small size and low-resolution. In SSPP scenario, how to accurately recognize the low-resolution faces is a more challenging task. Aiming at sovling the above problems, both the high-resolution and the low-resolution face recognition for single sample per person are investigated in this dessertation. In detail, the contents are as follows:
(1) Based on facial symmetry, a multiple feature subspace analysis (MFSA) approach is proposed for SSPP face recognition. In MFSA, the facial symmetry is used to augment the number of samples in the training set. First, MFSA divides each enrolled face into two halves about the bilateral symmetry axis, and further partition every half into several local face patches. Second, it clusters all the patches into multiple groups according to their locations at the half face and learns a feature subspace for each group of patches. By learning multiple feature subspace, the confusion between inter-class and intra-class variations of face patches is removed and more discriminative features can be extracted from each subspace. Compared with the state-of-the-art methods, MFSA is effortless and efficient in implementing, but achieves either better or competitive performance when recognizing face images taken in both constrained and unconstrained environment.
(2) Based on domain adaptation, an enhanced discriminative feature learning (EDFL) method is developed to deal with SSPP face recognition. EDFL is designed to make full use of existing face databases (named auxiliary dataset) which includes multiple face sample per person. EDFL treats the auxiliary dataset and gallery/probe dataset come from two domains (named source domain and target domain, respectively) with different data distributions, it is modeled from the perspective of domain adaptation. The aim of EDFL is to transfer the discriminative knowledge learned from the source domain to the target domain for classification. Firstly, considering that the galley set in target domain contains too limited sample to accurately represent the data distribution of the target domain, locality preserving domain adaption (LPDA) is presented to realize good overall domain adaptation. Secondly, to guarantee the learned domain adaptation components are optimal for discriminative learning, EDFL unifies the domain adaptation and discriminant model learning into the same framework. Meanwhile, to make the learned discriminant model robust to noises and outliers in the target domain, feature selection is taken into account in the model learning. EDFL is extensively evaluated under two scenarios, i.e., enhanced discriminative feature learning across ethnicity and across imaging condition. The comparison results with some state-of-the-art methods demonstrate the effectiveness of EDFL on SSPP face recognition.
(3) A cluster-based regularized simultaneous discriminant analysis (C-RSDA) method is developed to address low-resolution face recognition with SSPP. For low-resolution face recognition in SSPP scenario, it often suffers from overfitting and singular matrix problems. Because the clustering algorithms are able to gather similar samples while separate dissimilar ones, C-RSDA regularizes the between-class and within-class scatter matrices respectively with inter-cluster and intra-cluster scatter matrices, where the cluster-based scatter matrices are computed according to the clustering results of high-resolution face images. With the cluster-based scatter matrices, not only the singularity problem is resolved, but overfitting problem is overcomed as more variations are exploited from the limited training samples. C-RSDA is extensively evaluated on recognizing low-resolution face images captured in both controlled and uncontrolled environments. Compared with exiting methods designed for low-resoution face recognition, the encouraging experimental results of C-RSDA demonstrate its superiority over other methods.