Learning Effective Video Features for Facial Expression Recognition via Hybrid Deep Learning

Facial Expression Recognition is one of the recent trends to detect human expression in streaming video sequences. To identify emotions of video like sad, happy or angry. In this paper, the proposed method employs two individual deep convolution neural networks (CNNs), including a permanent CNN processing of static facial images and a temporary CN network processing of optical flow images, to separately learn high-level spatial and temporal characteristics on the separated video segments. Such two CNNs are fine tuned from a pre-trained CNN model to target video facial expression datasets. The spatial and temporal characteristics obtained at the segment level are then incorporated into a deep fusion network built with a model of deep belief network (DBN). This deep fusion network is used to learn spatiotemporal discriminative features together