Research on Polarity-Linear Attention-Based CNN Models for Emotion Recognition

This study proposes an improved CNN model incorporating Polar Linear Attention Mechanism for facial expression recognition tasks. Addressing the limitations of traditional CNNs in capturing long-range dependencies between facial expression regions, we achieve adaptive enhancement of key facial expression areas through deep integration of self-attention mechanisms with convolutional networks. Experiments on 7-class emotion classification using <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$64 \times 64$</tex>-pixel facial expression images compare the performance of standard CNN models, CNN with polar linear attention, and Transformer models. The results demonstrate that the CNN model with polar linear attention achieves an overall accuracy of 70.15%, representing a 1.44 percentage point improvement over the standard CNN model, while maintaining an optimal balance in inference speed and model size. The model exhibits particularly strong performance in recognizing expressions such as “happy” and “surprise”, validating the effectiveness of the proposed method in emotion recognition tasks.