doi:10.3850/978-981-08-7723-1_P235

ABSTRACT

We propose in this paper a novel enhanced subspace learning approach for facial expression recognition that is robust to misalignments. While subspace learning techniques have been successfully applied to appearance-based facial image analysis such as face recognition and facial expression recognition, most existing methods only work well when the face images are well aligned. In many real world applications such as visual surveillance and human robot interaction, there are usually some spatial misalignments in the cropped face images due to the eye localization error even if the eye positions are manually located, and the performance of appearance-based facial expression recognition methods degrade heavily under this scenario. To address this, we propose a novel enhanced subspace learning approach to generate more virtual misaligned samples to learn an enhanced feature subspace for facial expression recognition. Experimental results on two widely used face databases are presented to show the effectiveness and advantages of the proposed approach.

© AISM 2010 Organizers