基于不确定性感知学习的鲁棒表情识别方法

Uncertainty-Aware Learning for Robust Facial Expression Recognition Method

摘要: 真实场景下的面部表情识别(FER)深受数据模糊性和标签噪声的困扰。现有方法主要依赖确定性嵌入，将每张面部图像映射到特征空间中的一个固定点。然而，这种范式迫使模型将模糊样本(如存在遮挡、低分辨率或细微复合情绪的样本)过拟合到固定的类别标签上，进而降低了泛化能力。为解决这一问题，本文提出一种新颖的模糊感知与抑制模块(APSM)。与传统方法不同，APSM将面部特征建模为概率高斯分布，其特征由均值(语义中心)和方差(不确定性)表征。本文引入一种不确定性衰减损失函数(Uncertainty-Attenuated Loss)，该函数动态权衡学习过程：估计不确定性高的样本对梯度更新的贡献更小，从而有效抑制噪声数据的影响。在RAF-DB和AffectNet数据集上的大量实验表明，本文的方法在无需复杂外部数据或人工清理的情况下，显著提升了鲁棒性并达到了先进性能。

Abstract: Facial Expression Recognition (FER) in real-world scenarios is severely hampered by data ambiguity and label noise. Existing methods predominantly rely on deterministic embeddings, mapping each facial image to a fixed point in the feature space. However, this paradigm forces the model to overfit ambiguous samples—such as those with occlusion, low resolution, or subtle compound emotions—to rigid categorical labels, thereby degrading generalization capabilities. To address this, we propose a novel Ambiguity Perception & Suppression Module (APSM). Unlike traditional approaches, APSM models facial features as Probabilistic Gaussian Distributions, characterized by a mean (semantic center) and a variance (uncertainty). We introduce an Uncertainty-Attenuated Loss that dynamically weighs the learning process: samples with high estimated uncertainty contribute less to the gradient update, effectively suppressing the impact of noisy data. Extensive experiments on the RAF-DB and AffectNet datasets demonstrate that our method significantly improves robustness and achieves state-of-the-art performance without requiring complex external data or manual cleaning.