Face recognition with co-training and ensemble-driven learning

Face recognition using semi-supervised learning has received considerable amount of interest in the past years. In the same time, multiple classifier systems (MCS) have been widely successful in various pattern recognition applications such as fare recognition. MCS have also been very recently investigated in the context of semi-supervised learning. Very few attention has been devoted to verifying the usefulness of the newly developed semi-supervised MCS models for face recognition. In this work we attempt to access and compare the performance of several semi-supervised MCS training algorithms when applied to the face recognition problem. Experiments on a data set of face images are presented. Our experiments use non-homogenous classifier ensemble, majority voting rule and compare between various semi-supervised learning models. In particular we investigate the self-trained single classifier model, the ensemble driven model and a newly proposed modified co-training model. We also suggest the pair-wise classifier training strategy, which can be considered a special ensemble driven model. Experimental results reveal that the investigated semi-supervised models are successful in the exploitation of unlabelled data to enhance the classifier performance and their combined output. The proposed semi-supervised learning model based on co-training has shown a significant improvement of the classification accuracy compared to existing models.