抗噪声双约束网络的面部表情识别  

Facial Expression Recognition Based on Noise-Resistant Dual Constraint Network

作　　者：苏春海 夏海英[1,2] SU Chunhai;XIA Haiying(Guangxi Key Laboratory of Brain-inspired Computing and Intelligent Chips(Guangxi Normal University),Guilin Guangxi 541004,China;School of Electronic and Information Engineering/School of Integrated Circuits,Guangxi Normal University,Guilin Guangxi 541004,China)

机构地区：[1]广西类脑计算与智能芯片重点实验室(广西师范大学),广西桂林541004 [2]广西师范大学电子与信息工程学院/集成电路学院,广西桂林541004

出　　处：《广西师范大学学报(自然科学版)》2025年第2期70-82,共13页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金(62106054,62366006,62366005);广西创新驱动重大专项(桂科AA20302003);桂林市科技计划项目(20222C243986)。

摘　　要：由于标注主观性、图像模糊等因素,数据集不可避免存在噪声,使表情识别更具挑战性。现有面部表情识别方法在处理噪声标签时,模型会部分过度拟合噪声标签,对此,本文提出一种新颖的抗噪声双约束网络(NDC-Net)来自动抑制噪声样本。NDC-Net主要包括2个约束机制:类激活映射注意一致性(CAC)和通道空间特征一致性(CSC)。CAC使模型集中于局部重要特征信息,减少对噪声标签的过度关注,而CSC鼓励和确保模型在提取特征时从通道和空间上更加关注到与任务相关的信息,忽略不相关信息,减少对噪声标签的依赖。此外,为增强NDC-Net性能,输入样本采用旋转、缩放等策略进行增强。在RAF-DB、FERPlus和AffectNet数据集30%标签噪声下,NDC-Net的识别性能分别为86.57%、88.22%和59.78%,显著优于EAC、NCCTFER等先进的噪声标签处理方法,并且在计算机视觉领域中被广泛应用于评估算法性能和泛化能力的CIFAR100和Tiny-ImageNet中也取得不错的效果。Noise is inevitably present in datasets due to labeling subjectivity,image blurring,and other factors,making expression recognition more challenging.Existing facial expression recognition methods typically address noisy labels by partially overfitting to them.In this paper,a novel Noise-Resistant Dual Constraint Network(NDC-Net)is proposed to automatically suppress noisy samples.NDC-Net primarily consists of two constraint mechanisms:Class Activation mapping attention Consistency(CAC)and Channel and Spatial feature Consistency(CSC).CAC is used to make the model focus on locally important feature information and reduces the overfitting to noisy labels,while CSC is used to ensure that the model emphasizes task-relevant information from both channels and spatial dimensions during feature extraction,ignoring irrelevant information,and reducing reliance on noisy labels.Additionally,to enhance the performance of NDC-Net,input samples are augmented with strategies such as rotation and scaling.NDC-Net achieves recognition performances of 86.57%,88.22%,and 59.78%under 30%label noise for RAF-DB,FERPlus,and AffectNet datasets,respectively.These results significantly outperform the state-of-the-art noisy labeling methods,such as EAC,NCCTFER.Moreover,NDC-Net also shows strong generalisation capability on general classification datasets such as CIFAR100 and Tiny-ImageNet.

关 键 词：噪声标签 面部表情识别 深度学习 监督学习 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]