Gradient-Based Compact Binary Coding for Facial Expression Recognition

Abstract

Facial Expressions are one of the most essential and general means for human beings to express their feelings and communicate their emotions; thus, Facial Expression Recognition (FER) has gained a significant research interest and an attractive topic in Human-Computer Interactions (HCI). Towards such, this paper proposes a novel face descriptor, Gradient Direction Pattern (GDP), for facial expression recognition. With the help of gradients, GDP encodes the structure of the facial image in a more compact way such that the FER is robust to pose variations. Furthermore, the GDPs are measured for edge feature maps, extracted from Gaussian filtered and Gabor filtered facial images at different orientations. These edge feature maps make the recognition system robust to noise, illumination, scaling, and orientational variations. Initially, the facial image is divided into small regions, and GDPs are extracted from them. Then these patterns are concatenated, and Histograms are measured, called GDP Histograms (GH). Simulation experiments were conducted over two standard datasets, such as CK+ and JAFFE, and the average accuracy is observed as 95% and 91% approximately.