Real-World Emotion Recognition Using Edge AI

The education sphere, healthcare, automotive safety, social robots, and smart technology sectors require emotion recognition through speech and facial expressions. Nonetheless, deep learning technologies are based on cloud computing which results in longer response times and needs data storage that fits sensitive data for transfer. With Edge AI, it becomes possible to perform inference in the proximity of the sensing device. This paper aims to explore lightweight architecture types and implementation techniques implementing emotion recognition at the edge. The types of architecture and techniques discussed throughout the paper include small CNNs, transformer-based models, and deep learning model compression methods, as quantization, pruning, knowledge engineering, and NAS. The article discusses a comparison between the types of implementations used for the technologies (embedded hardware and inference frameworks) and forms of optimization techniques. In addition, it compares the various studies that have been conducted on them regarding their implementation, type of data sets, and effectiveness, including the ones that failed to mention the latency of their system.