Emotion Recognition via Environmental Context and Human Body

To promote the humanized interactive experience of the intelligent device and system, emotional intelligence has become a popular research field in the human-machine interaction. The previous research on emotion recognition based on computer vision has mostly been carried out by analysing facial expression or body posture, and psychological studies show that scene context also contributes some important information on emotion recognition. In addition, most context-aware emotion recognition studies focus on exploring the relevance analysis of environmental semantics, but the influence of feature encoder on semantic information embedding has not been fully discussed. In this paper, we proposed a Global Semantic Feature Enhancement-Dual Stream Densely Connected Network (GSFE-DSDCN) to enhance global semantic information learning from the perspectives of dimension and spatial. Densely connected pattern is introduced to concatenate the shallow and deep layers output, which fuses the semantic information of low-dimensional geometric features and high-dimensional abstract context features together. The Global Multi-Scale Feature Recalibration (GMSFR) module expands the receptive field in spatial, which effectively improves the global semantic features extraction capability of feature encoder. We evaluate the proposed method on the EMOTIC data set, and experimental results are shown to be competitive with the state-of-the-art algorithms.