A Research On Feature Extraction And Recognition Of Occluded Face Images

As face recognition technology has the characteristics of safety, secrecy, friendly and convenience, it has widely been used in finance, information assurance and public security since1960 s. However, real-world face images are often partially occluded. The complexity and unpredictability of partial occlusion pose a great challenge for face recognition. Thus, what is critical for face recognition with occlusion is to design an efficient method to weaken the influence incurred by occlusion. Recently, many researchers made a thorough study on this issue,but there still exits some problems: 1) The algorithms are not robust to high-level occlusions and usually have a breakdown point beyond which they are sensitive to occlusions and their recognition performances drop sharply. 2) Other variations except for occlusion might simultaneously exist, such as small sample size of the training images or illumination variations in the test images.In order to remedy the above problems, we propose a Bayesian multi-distribution-based discriminative feature extraction method and a robust face recognition method by fusing gradientface with Markov random fields. The main contributions are as follows:(1) In order to solve the small sample size problem, we propose a Bayesian multi-distribution-based discriminative feature extraction method. To enlarge the number of the training samples of each subject, we first divide each sample into several patches. It also helps to scatter the occlusion, and thus reduce the effect of occlusion on recognition. We then propose a Bayesian learning framework to extract the discriminative features from the divided patches.Specifically, we transform the features of the intra-class patches into a new low-dimensional subspace by maximizing a multivariate Gaussian likelihood function, and, simultaneously,enlarge the distances between the inter-class features by maximizing an exponential priori distribution. For classification, we use the nearest neighbor classifier combined with theMahalanobis distance, instead of Euclidean distance.(2) In order to solve the problem with high-level occlusion and illumination variations, we propose a robust face recognition method by fusing gradientface with Markov random fields.Gradientface is usually utilized as features against illumination variations. In the domain of gradientface, we build the probabilistic generative model of the reconstruction error, which combines the distribution of the reconstruction error conditioned on the occlusion support and the priori probabilistic distribution of the occlusion support. Specifically, in order to discriminate the occluded region from the non-occluded region, we build the conditional distributions of the reconstruction error in the two regions, separately. That is, we build the error distribution in the occluded region with the uniform distribution and the one in the non-occluded region with the Gaussian distribution. Meanwhile, we also consider the spatial contiguous prior of the occlusion,which can be modeled by a Markov Random Field. This method can locate the occlusion accurately. Then we can use the non-occluded feature to improve the recognition performance.Experiments on two widely-used face databases verify the effectiveness of the proposed methods. In future work, we will consider more variations, such as expression and pose variations, in our error probabilistic model to improve its robustness in real-world face recognition.