基于特征融合与多注意力机制的人脸表情识别方法

Facial Expression Recognition Method Based on Feature Fusion and Multi-Attention

摘要: 当前，人脸表情识别(FER)领域正受到越来越多研究者的关注。然而，FER仍面临诸多挑战，特别是在真实场景(in-the-wild)数据集上，模型性能的提升受到显著限制。为了解决FER领域中身份信息干扰、类内差异和类间相似等问题，本文提出了一种基于特征融合与多注意力机制的人脸表情识别方法。该方法首先采用金字塔结构整合骨干网络提取的多层语义特征。然后，通过多注意力模块来强化模型全局与局部的学习能力，自适应地捕捉关键的人脸信息。最后，使用正交分解映射对融合特征进行分解，进一步去除特征间的冗余信息，尤其是身份干扰信息，从而获得具有高辨别性以及良好泛化能力的表情特征。我们所提出方法在RAF-DB、AffectNet数据集上实现了优秀的效果，大量实验证明了方法的有效性。

Abstract: Currently, the field of Facial Expression Recognition (FER) is attracting increasing attention from researchers. However, FER still faces numerous challenges, particularly in enhancing model performance on real-world (in-the-wild) datasets. To address issues such as identity information interference, intra-class variation, and inter-class similarity in FER, this paper proposes a facial expression recognition method based on feature fusion and a multi-attention mechanism. The method first employs a pyramid structure to integrate multi-layer semantic features extracted by the backbone network. Then, it uses multi-attention modules to enhance the model’s ability to learn both global and local features, allowing for adaptive capture of key facial information. Finally, orthogonal decomposition mapping is applied to the fused features to further eliminate redundant information between features, especially identity interference, resulting in expression features with high discriminability and good generalization ability. Our proposed method achieves excellent results on the RAF-DB and AffectNet datasets, with extensive experiments demonstrating its effectiveness.