Spontaneous Expression Recognition Based on Visual Attention Mechanism and Co-salient Features

摘要

Spontaneous facial expression recognition has gained much attention from researchers in recent years, however most of the existing algorithms still encounter bottlenecks in performance due to too big redundant images data in the video. In this paper, we propose a novel co-salient facial feature extraction algorithm, combined with human visual attention mechanism and group data coprocessing technology, which would largely reduce the redundant information in the original images and effectively improve the recognizing accuracy of facial expressions. Firstly, based on human visual mechanism, key frames of expression are dynamically derived from the original videos to capture the temporal dynamics of facial expressions. Secondly, using key sequence frames, salient regions are obtained by multiplicative fusion algorithm and in multi-images co-operative manner. Thirdly, we get rid of these salient regions due to their little deformation and low-correlation to facial expressions, and reduce the number of facial features data. At last, we extract Local Binary Pattern (LBP) features from the remainder of facial features and use Support Vector Machine (SVM) classifier to classify them respectively. Experimental results on dataset Cohn-Kanade plus and MMI showed that our proposed method can effectively improve the recognizing accuracy of spontaneous expression sequence.