Abstract
Facial expression recognition (FER) is a crucial task in computer vision with wide range of applications including human computer interaction, surveillance, and assistive technologies. However, challenges such as occlusion, expression variability, and lack of interpretability hinder the performance of traditional FER systems. Graph Neural Networks (GNNs) offer a powerful alternative by modeling relational dependencies between facial landmarks, enabling structured and interpretable learning. In this paper, we propose GLaRE, a novel Graph-based Landmark Region Embedding network for emotion recognition. Facial landmarks are extracted using 3D facial alignment, and a quotient graph is constructed via hierarchical coarsening to preserve spatial structure while reducing complexity. Our method achieves 64.89 percentage accuracy on AffectNet and 94.24 percentage on FERG, outperforming several existing baselines. Additionally, ablation studies have demonstrated that region-level embeddings from quotient graphs have contributed to improved prediction performance.
Abstract (translated)
面部表情识别（FER）是计算机视觉中的一个关键任务，具有广泛的应用场景，包括人机交互、监控和辅助技术。然而，遮挡、表情变化多样性和缺乏可解释性等问题限制了传统FER系统的性能。图神经网络（GNNs）通过建模面部特征点之间的关系依赖性提供了一个强大的替代方案，这使得结构化且易于理解的学习成为可能。在本文中，我们提出了GLaRE，这是一种新型的基于图的地标区域嵌入网络，用于情感识别。使用三维面部对齐提取面部关键点，并通过分层粗化构建商图以保持空间结构的同时减少复杂性。我们的方法在AffectNet数据集上实现了64.89%的准确率，在FERG数据集上达到了94.24%，优于多个现有基线模型。此外，消融研究证明了来自商图的区域级嵌入对预测性能有所改善。
URL
https://arxiv.org/abs/2508.20579
