Research on Facial Expression Recognition Based on Multi-head Attention Mechanism

Facial expressions are among the most natural and efficient means of communicating emotions and intentions, accounting for around 55% of human communication when expressing various emotional attitudes. On the other hand, facial expressions are inherently complicated, varied, and comparable. Facial expression recognition (FER) is made more difficult in natural settings by extra difficulties such partial occlusions, lighting variances, and individual differences. We suggest a unique Multi-head Attention Affinity Diversity Sharing Model (MADS) based on attention mechanisms to overcome these obstacles and increase the precision of facial emotion identification in natural settings. We present a sparse affinity loss function in our approach. We use attention weights to highlight the most discriminative features because typical affinity loss functions are vulnerable to redundant and noisy information because they equally contribute to the objective across all feature dimensions. We also suggest a diversity loss function to improve the multi-head attention network’s guidance. The proposed framework employs a diversity-promoting objective function that directs attention to multiple distinct facial areas while simultaneously maximizing the differentiation between spatial and channel attention patterns. Additionally, we implement a shared adaptive feature integration mechanism that dynamically combines information across all attention heads. Given the high intrinsic similarity among facial expressions, we decompose expression features into a combination of generic and unique features. The sharing mechanism allows the network to learn differences based on generic features rather than relearning complete features from scratch. Comprehensive evaluations conducted on four benchmark naturalistic facial expression datasets confirm both the efficacy and generalization capability of our approach. The experimental results reveal that MADS consistently surpasses existing state-of-the-art methods, establishing new accuracy benchmarks across all evaluated datasets. Further analysis of the confusion matrix, precision-recall curve, and computational complexity further confirms the advantages of our model.