EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Dual‐EmoNet: An Efficient Network for Fine‐Grained Facial Expression Recognition With Dual Attention

Weihao Zeng, Wei Liu

IET Image Processing · 2026-01 · 20 (1)

DOI: 10.1049/ipr2.70301 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IET Image Processing, online erschienen am 09.02.2026. Verfasst von Weihao Zeng und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Natural Science Foundation of Hunan Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The practical application of facial expression recognition (FER) is significantly hampered by three key challenges: insufficient discrimination between fine‐grained expressions (e.g., “sad” vs. “disgust”), high computational costs of state‐of‐the‐art (SOTA) models, and the inherent class imbalance in datasets. These challenges severely constrain the reliability of FER systems and their deployment in scenarios requiring precise emotional understanding. To address these challenges, this study proposes Dual‐EmoNet, a novel and efficient deep learning architecture for fine‐grained FER. Dual‐EmoNet enhances discriminative features by integrating an innovative Dual‐ResBlock with a dual‐attention mechanism (synergising channel and spatial attention), reduces computational overhead through lightweight GhostNorm convolutions, and combines a pyramid pooling module with an ArcFace‐based classification head to synergistically improve both fine‐grained recognition and computational efficiency. Dual‐EmoNet was comprehensively evaluated using multiple FER datasets, including the Real‐world Affective Face Database (RAF‐DB), FERPlus, and AffectNet. Experimental results demonstrate that Dual‐EmoNet achieves a best test accuracy of 95.37%, with an overall accuracy of 95.09% ± 0.21%, significantly outperforming previous SOTA methods such as QCS (93.02%) and POSTER++ (92.21%), while maintaining a low parameter count of 42.49 M and just 0.46 GFLOPs (giga floating‐point operations). The proposed model further attains competitive performance on FERPlus (93.13% ± 0.19%) and AffectNet‐7 (67.03% ± 0.15%), demonstrating strong robustness under both controlled and in‐the‐wild conditions. Cross‐dataset evaluations reveal favourable generalisation ability, particularly when trained on large‐scale AffectNet‐7 and tested on RAF‐DB, highlighting the effectiveness of learning transferable expression representations. This work demonstrates that it is possible to effectively address the inherent trade‐offs in FER – granularity, efficiency, and data imbalance – through synergistic multi‐component optimisation, thus providing a robust solution for developing high‐performance and easily deployable affective computing systems for resource‐constrained environments. The source code and trained model will be made available at https://github.com/romanceyu/Dual‐EmoNet .

Förderung

  • Natural Science Foundation of Hunan Province

Zitieren

APA

Zeng, W., & Liu, W. (2026). Dual‐EmoNet: An Efficient Network for Fine‐Grained Facial Expression Recognition With Dual Attention. IET Image Processing, 20(1). https://doi.org/10.1049/ipr2.70301

Vancouver

Zeng W, Liu W. Dual‐EmoNet: An Efficient Network for Fine‐Grained Facial Expression Recognition With Dual Attention. IET Image Processing. 2026;20(1). doi:10.1049/ipr2.70301

BibTeX
@article{Zeng2026,
  author = {Zeng, Weihao and Liu, Wei},
  title = {Dual‐EmoNet: An Efficient Network for Fine‐Grained Facial Expression Recognition With Dual Attention},
  journal = {IET Image Processing},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/ipr2.70301}
}
Kostenlos starten →