Facial Expression Recognition Based on Group Domain Random Frame Extraction

摘要

Modeling the dynamic variation of facial expression from a sequence of images is a key issue in facial expression recognition. However, the analysis of complete sequence temporal information requires significantly computational power. To improve the efficiency, a dynamic frame sequence convolutional network (DFSCN) is proposed in this study. In the proposed DFSCN, an expression sequence simplification method is first proposed to reduce the sequence length and takes the reduced new sequence as the input of DFSCN. An adaptive weighted feature fusion method for spatiotemporal feature learning is then put forward in DFSCN. A still frame convolutional network (SFCN) is introduced for complementing the still appearance information and the fine-tuning of DFSCN. Finally, these two models are combined together by weighted fusion to enhance the performance. Two public-available databases, CK+ and Oulu-CASIA, are used to evaluate the performance of the proposed approach. Experimental results show that the proposed method can effectively capture the dynamic process of expression sequence and the recognition performance is superior to other state-of-the-art methods.