Research on Video Emotion Recognition Method Based on Deep Learning

Abstract

Video data has significant temporal and spatial characteristics, and the extraction and recognition of its emotional features is currently a research hotspot in the field of human-computer interaction. This paper proposes an improved model based on P3D 3D residual network (3Res att Network) to address the difficulties in dynamic feature extraction and insufficient spatiotemporal information fusion in video emotion recognition tasks. Firstly, the P3D infrastructure is constructed by decoupling spatiotemporal convolution kernels, effectively reducing model complexity; Secondly, design a 3D Spatial Attention mechanism to dynamically focus on emotionally significant areas and enhance feature discriminability; Finally, by combining residual connections with multi head attention mechanism, an end-to-end 3Res att network is constructed to achieve deep fusion of spatiotemporal features. Experiments on the eNTERFACE '05 dataset show that our method achieves an accuracy of 67.5% in video emotion recognition tasks, which is 18.2% higher than traditional C3D models and 11.4% higher than 3Res att2. The ablation experiments further validated the effectiveness of the attention module, providing a new technological path for video emotion computing.