Circuit Design of Multimodal Attention Memristive Network for Affective Video Content Analysis

Abstract

Affective video content analysis aims at automatically identifying human emotion triggered by video, which plays an important role in mental health monitoring. This paper proposes a multimodal attention memristive network for affective video content analysis, which offers an energy-efficient approach with low time consumption and high classification accuracy. To illustrate the complexity of the proposed multimodal attention memristive network, two core modules are proposed. Firstly, unimodal feature representation module with cascaded configuration is designed to capture unique characteristics from multimodal signals. Then, multimodal local-global fusion module is proposed to stimulate the process of multimodal information sensing and processing in human brain. Furthermore, the proposed system is validated by applying it to affective content analysis. The experimental results demonstrate that the multimodal attention memristive network outperforms the existing state-of-the-art methods with high classification accuracy and low time consumption.