Modular PCA and probabilistic similarity measure for robust face recognition

This paper addresses a probabilistic approach to develop a robust face recognition system to partial variations such as occlusions. Based on the statistical feature extraction methods, we take the modular PCA method which finds eigenspace not for the set of whole images but for the sets of local image patches. Through the local feature extraction approach, we try to overcome the drawback of wholistic appearance-based conventional PCA, and consequently expect to improve robustness to local variations. The obtained local features are then applied to define two probabilistic models for facial images: one for modeling distribution of features observed in usual facial images, and the other for modeling distribution of environmental variations observed in face image from one subject. The probabilistic model for general facial images are used to evaluate the importance of each local patch. The probabilistic model for the environmental variations is used to evaluate the similarity between two local features. By combining two probabilistic models, we finally define a distance measure between two face images, which can be applied for face recognition. Computational experiments on benchmark face database show that the proposed face recognition method can achieve remarkable robustness to local variations.