Research on Improved Facial Emotion Recognition Model with Multi-Head Attention Mechanism in Lightweight Environment

Although facial emotion recognition technology has made remarkable achievements, it still faces key challenges such as long model training times and high model complexity. Therefore, this dissertation will focus on analyzing the application of lightweight models for human facial emotion recognition, as well as investigating the improvement of adding attention to lightweight models and exploring possible solutions. The experimental section comprises three sets of evaluations. Firstly, lightweight vs. heavyweight models are contrasted. Secondly, the efficiency of three lightweight models is scrutinized. Finally, the addition of an attention mechanism to a lightweight model is tested for enhanced emotion recognition. This comparison reveals the chosen lightweight model’s superiority over heavyweight alternatives in terms of lightness and performance. It also assesses the efficacy of different neural architectures in facial emotion recognition. In conclusion, MobileNetV2 emerged as the top performer, with its accuracy marginally improved to ${6 7.6 4 \%}$ after integrating the designed attention mechanism in 200 training epochs. This finding offers crucial insights for future applications in resource-constrained settings.