Effective curvelet-based facial expression recognition using graph signal processing

Abstract

Features in the facial images are multi-dimensional. Different facial expressions-based images are actually the interplay of the edges, which are found on faces. The wavelet transform which has been extensively used as a tool for mathematical analysis of the facial images has the disadvantage of poor directionality. Thus, curvelet transform is preferred for the facial image analysis due to better representation of edges by the directional elements. However, its feature dimension is of large size which makes the curvelet approach computationally expensive. In order to capture the interrelationship in the curvelet transform-based features at the higher level to construct new feature vectors, the proposed work suggests the use of graph signal processing along with the curvelet transform for recognizing the facial expressions. Not only the dimension of the feature vectors has been reduced but also recognition of the facial expression has been significantly improved. Experiments for Japanese female facial expression database and Cohn–Kanade (CK+) database show the effectiveness of the proposed approach.