SHCFNet on Micro-expression Recognition System

Micro expression is a facial feature that can reflect the most real emotional state hidden in the human heart. This is a very short process and difficult to capture accurately. Based convolutional network, a new network architecture (SHCFNet) is proposed to extract the spatial-temporal feature of peak frames, the optical flow between onset and apex frame and its derivative (optical strain). The proposed network stacks these features from the outcomes of the previous layer. Then, the stacked feature is merged with the convolution feature of the previous layer, which enhances the learnability of neurons. The performance of the proposed SHCFNet are evaluated on four benchmark datasets: CASME I, CASME II, SAMM and SMIC, and compared with other advanced networks.