A co-training approach to automatic face recognition

Semi-supervised face recognition using both labelled and unlabelled data has received considerable interest in recent years. Co-training is one of the most well-known semi-supervised learning methods, but its application in face recognition almost remains unexplored because its assumption of conditional independence can be rarely satisfied between two facial features. However, even if two facial features are not completely independent, their different characteristics produce a so-called “classification margin” between two classifiers based on them, and hence there is the possibility of mutual training. In this paper, we report a semi-supervised face recognition algorithm which applies co-training on two classifiers based on Linear Discriminant Analysis (LDA) and Local Binary Patterns (LBP) features respectively. Experimental results show not only that the proposed co-training algorithm significantly improves the recognition accuracy over supervised methods using only labelled training data, but also demonstrates the superiority of co-training over self-training methods which only use one facial feature.