Facial expression recognition based on multi-scale CNNS

Abstract

This paper proposes a new method for facial expression recognition, called multi-scale CNNs. It consists several sub-CNNs with different scales of input images. The sub-CNNs of multi-scale CNNs are benefited from various scaled input images to learn the optimalized parameters. After trained all these sub-CNNs separately, we can predict the facial expression of an image by extracting its features from the last fully connected layer of sub-CNNs in different scales and mapping the averaged features to the final classification probability. Multi-scale CNNs can classify facial expression more accurately than any single scale sub-CNN. On Facial Expression Recognition 2013 database, multi-scale CNNs achieved an accuracy of 71.80% on the testing set, which is comparative to other state-of-the-art methods.