Pose Overcomplete Automatic Registration Method for Video Based Robust Face Recognition

Abstract

In ideal conditions with good illumination, posevariation, resolution, and registration, the performance ofrecognition system has reached a relatively high level. However,the impact of pose variation and registration factors hasntbeen well solved. In practical system, single face with manuallyregistration is often adopted in the registration process. Butthere are some limitations in the single registered face image,and manually way of registration is not convenient for theusers. Also the recognition performance based on the singleface image matching is inevitably disturbed and restricted. Inpractical situation, more than one face images can be collected.Multiple image of the same person than a single image featurecan capture more intra variation information in the same class.Feature information of multiple images are introduced in therecognition matching, which will benefit to improve theaccuracy of face recognition. In this paper, a poseovercomplete automatic registration method is proposed tosolve this problem in the registration process. In the proposedmethod, we estimate the pose automatically in real-time byutilizing the detected landmarks information, with posevariation face images stored for matching templates.Experimental results show that the proposed method can notonly overcome the influence of the pose variation inrecognition, but also can solve the problem of non-ideal poseregistration, thus improve the recognition accuracy in practicalface recognition system.