An Incremental Learning Face Recognition System Based On Subspace Method

The variability of face images under uncontrolled conditions and the difficulties of collecting samples may be the most important two challenges in face recognition applications. Though numerous recognition algorithms have been proposed in the past twenty years, most of them failed under the condition of one training sample per per-son, due to the inherent complexity of face images and the the lack of training data.In this paper, we propose a novel incremental learning face recognition system based on subspace method.To deal with the Small Sample Size(SSS) Problem, each face images is presented as a series of subspaces by the sample set enlarging. These subspaces are presented as a fixed number of orthonormal bases. The recognition problem can be addressed by measuring the similarity between the unlabeled subspaces and the training sample subspaces.Inspired by the concept of incremental learning and semi-supervised learning, we try to employ the inputting data to dramatically improve the performance of the recognition system. By proposing a novel subspace updating strategy, the training sample subspaces can be automatically adjusted and updated with new information extracted from inputting unlabeled face data with acceptable additional computational cost.To realize robust face recognition, we design a subspace similarity measure based on principal angles which works well in our task. We also propose a threshold policy which can avoid learning distorted information during the incremental learning pro-cess. Finally, experiments on well-known face databases such as AR and Extended YALE are taken. And the results show that our subspace updating strategy can sig-nificantly improve the recognition accuracy. Moreover, our proposed approach outper-forms several existing preprocessors in the scenario of one training sample per person with significant facial variations.