Deep spatial-temporal feature fusion for facial expression recognition in static images

Traditional methods of performing facial expression recognition commonly use hand-crafted spatial features. This paper proposes a multi-channel deep neural network that learns and fuses the spatial-temporal features for recognizing facial expressions in static images. The essential idea of this method is to extract optical flow from the changes between the peak expression face image (emotional-face) and the neutral face image (neutral-face) as the temporal information of a certain facial expression, and use the gray-level image of emotional-face as the spatial information. A Multi-channel Deep Spatial-Temporal feature Fusion neural Network (MDSTFN) is presented to perform the deep spatial-temporal feature extraction and fusion from static images. Each channel of the proposed method is fine-tuned from a pre-trained deep convolutional neural networks (CNN) instead of training a new CNN from scratch. In addition, average-face is used as a substitute for neutral-face in real-world applications. Extensive experiments are conducted to evaluate the proposed method on benchmarks databases including CK+, MMI, and RaFD. The results show that the optical flow information from emotional-face and neutral-face is a useful complement to spatial feature and can effectively improve the performance of facial expression recognition from static images. Compared with state-of-the-art methods, the proposed method can achieve better recognition accuracy, with rates of 98.38% on the CK+ database, 99.17% on the RaFD database, and 99.59% on the MMI database, respectively.