Adaptively weighted facial expression recognition by feature fusion under intense illumination condition

Abstract

Accurate and robust facial expression recognition under complex environment is a challenging task. In this paper, we propose an adaptively weighted facial expression recognition approach to overcome the intense illumination difficulty by fusing diverse illumination invariant appearance features. First, a novel neural-network-based adaptive weight assignment strategy is designed to eliminate the adverse illumination variations efficiently and effectively. Then, a feature fusion strategy is developed to combine two of the most successful illumination invariant appearance descriptors, namely Gabor and Local Binary Patterns (LBP), for giving comprehensive and robust description of facial expressions. Extensive experiments demonstrate the superiority of the proposed approach on the common used CK+ dataset, especially the adaptive weight assignment for the significant improvement of recognition accuracy under extreme and intense illumination conditions.