Fuzzy Membership-Driven Adaptive Sample Mining for Facial Expression Recognition
IEEE Transactions on Fuzzy Systems · 2025-12 · 33 (12) · 4240-4251
DOI: 10.1109/tfuzz.2025.3593951 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Fuzzy Systems, 2025-12. Verfasst von Xuxiong Liu und 4 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 64 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Anhui Province Key R&D Program, General Programmer of the National Natural Science Foundation of China.
Abstract (Originalsprache)
Facial expression recognition (FER) datasets are plagued by fuzzy label uncertainty due to interclass similarities, compound emotions, and annotation subjectivity. Moreover, the complexity of recognition differs across various expression categories, making an undifferentiated method inequitable for each expression. In this work, we propose adaptive sample mining (ASM), a fuzzy membership-driven framework with two innovations: it constructs a category-specific dynamic threshold set as the average probabilities of samples that are correctly predicted, through which it explicitly separates clean and noisy subsets while leaving ambiguous samples as the intermediate fuzzy region between them. In addition, it introduces a rule based on multilocal region prediction consistency to reclassify samples from the clean subset into the ambiguous subset when local predictions conflict. Finally, the triregularization module implements tailored training approaches for these three subsets: the clean subset is trained using a supervised learning strategy to maximize the utility of high-quality annotations, the ambiguous subset adopts a mutual learning strategy to enhance discrimination capability, and the noisy subset employs an unsupervised learning technique so that it can minimize the influence from noisy labels. We performed comprehensive experiments on the RAF-DB, FERPlus, and AffectNet. The proposed ASM achieved 90.94% accuracy on RAF-DB and 86.81% on its synthetic noisy version. The ablation experiment and analysis demonstrate that our method can efficiently identify ambiguous and noisy labels. By weighting the classification loss based on fuzzy membership, it helps to make efficient use of the labels. This results in an improvement in accuracy.
Förderung
- National Natural Science Foundation of China
- Anhui Province Key R&D Program
- General Programmer of the National Natural Science Foundation of China
Zitieren
APA
Liu, X., An, L., Shen, K., Zhang, Z., & Sun, X. (2025). Fuzzy Membership-Driven Adaptive Sample Mining for Facial Expression Recognition. IEEE Transactions on Fuzzy Systems, 33(12), 4240-4251. https://doi.org/10.1109/tfuzz.2025.3593951
Vancouver
Liu X, An L, Shen K, Zhang Z, Sun X. Fuzzy Membership-Driven Adaptive Sample Mining for Facial Expression Recognition. IEEE Transactions on Fuzzy Systems. 2025;33(12):4240-4251. doi:10.1109/tfuzz.2025.3593951
BibTeX
@article{Liu2025,
author = {Liu, Xuxiong and An, Liuwei and Shen, Kang and Zhang, Ziyang and Sun, Xiao},
title = {Fuzzy Membership-Driven Adaptive Sample Mining for Facial Expression Recognition},
journal = {IEEE Transactions on Fuzzy Systems},
year = {2025},
volume = {33},
number = {12},
pages = {4240--4251},
doi = {10.1109/tfuzz.2025.3593951}
}
EUVIMED