Multi-view semi-supervised discriminant analysis: A new approach to audio-visual person recognition

Many state-of-the-art biometric systems use feature vectors of high dimension and call for dimensionality reduction techniques to avoid the co-called `curse of dimensionality.´ Supervised approaches such as Linear Discriminant Analysis can extract discriminative features and is used widely, but suffers from over-fitting when used with small datasets. Through the imposition of local adjacency constraints, semi-supervised dimensionality reduction techniques can make use of abundant, unlabelled data to improve classification performance. This paper reports a new multi-view, semi-supervised discriminant analysis (MSDA) algorithm and its application in audio-visual person recognition. In contrast to existing approaches which typically utilize a single view, MSDA determines a more reliable neighbourhood constraint built jointly from multiple views of the same data. Experimental results on the standard MOBIO database show that our algorithm not only outperforms baseline supervised and unsupervised methods, but that it also outperforms single-view semi-supervised dimension reduction techniques in single view.