A method for facial expression recognition of image sequence based on spatial features

Abstract

Facial expression recognition is the foundation of human emotion recognition, which has become a hot topic in the field of artificial intelligence in recent years. To some extent, facial expressions can be regarded as the process of facial muscle changes. Image sequence contains richer expression contents compared with a single image. So the expression recognition based on image sequence can yield more accurate results. A new method for facial expression recognition is proposed in this paper, which is based on spatial feature for image sequence. The method consists of four steps. Firstly, Siamese neural network is used to construct an evaluation model for changes of expression intensity, which extracts an appropriate image sequence from the video. Secondly, a convolutional neural network with attention mechanism is designed and trained, which is used to extract spatial features from each image in the sequence. Then, the spatial features of multiple images are fused. Finally, the fusion results are put into a convolutional neural network to recognize the facial expressions. This method is validated on CK+ dataset and the experimental results show that it’s more accurate than several other methods.