A Graph Attention Network-Enhanced Approach to Facial Expression Recognition Using Hybrid Pixel-Geometry Features

Facial Expression Recognition (FER) plays a pivotal role in various fields such as human-computer interaction, psychological assessment, and education, aiming to interpret emotions through facial expressions.However, FER faces significant challenges due to environmental variations like lighting, head poses, and occlusions.Appearance-based methods are easily affected by lighting and face angles, while geometry-based methods struggle when the face is occluded or head pose changes.Both lack flexibility in handling real-world complexities, reducing facial expression recognition performance.In this paper, we propose a novel hybrid approach that integrates Graph Attention Networks (GAT) for geometry-aware feature extraction and Residual Masking Networks (RMN) for pixellevel feature refinement.This dual-stream model captures both the spatial relationships among facial landmarks and subtle facial texture variations, allowing it to more accurately distinguish emotions.Attention mechanisms are employed to prioritize key regions of the face, improving the model's robustness and precision.Experimental results on benchmark datasets, including CK+, JAFFE, RAF-DB and FER2013 demonstrate the superiority of our approach.Experimental evaluation on four standard benchmarks demonstrates the efficacy of our approach: it attains a perfect 100% classification accuracy on the six emotion categories of CK+, 95.8% on JAFFE, 83.8% on the more challenging RAF-DB, and 74.9% on the in-the-wild FER2013 dataset.Moreover, our approach demonstrates competitive performance relative to current state-of-the-art FER methods across these challenging benchmarks.The proposed method effectively combines the strengths of appearance and geometric feature analysis, leading to a more robust and reliable solution for facial expression recognition.