Dynamic visual time context descriptors for automatic human expression classification

Abstract

In this paper, we propose two fast dynamic descriptors Vertical-Time- Backward (VTB) and Vertical-Time-Forward (VTF) on spatial-temporal domain to catch the cues of essential facial movements. These dynamic descriptors are used in a two-step system to recognize human facial expression within image sequences. In the first step, the system classifies static images and then it identifies the whole sequence. After combining the visual-time context features with popular LBP, the system can efficiently recognize the expression in a single image, and is especially helpful in highly ambiguous ones. In the second step, we use the evaluation method through the weighted probabilities of all frames to predict the class of the whole sequence. The experiments were performed on 348 sequences from 95 subjects in Cohn-Kanade database and obtained good results as high as 97.6 % in seven-class recognition for frames and 95.7 % in six class for sequences. © Springer-Verlag Berlin Heidelberg 2014.