Contrastive Decoupling: Dynamic Regularization for Enhanced Fine-Grained Image Classification
IEEE Transactions on Circuits and Systems for Video Technology · 2026-06 · 36 (6) · 7588-7599
DOI: 10.1109/tcsvt.2026.3661760 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Circuits and Systems for Video Technology, 2026-06. Verfasst von Zheyuan Wang und 3 weiteren Autor:innen. Die Arbeit wurde bisher 31-mal zitiert (Crossref) und nennt 97 Literaturangaben. Gefördert von: National Natural Science Foundation of China NSFC, Noncommunicable Chronic Diseases–National Science and Technology Major Project, National Key Research and Development Program of China.
Abstract (Originalsprache)
Fine-grained visual classification remains challenging due to subtle inter-class differences and significant intra-class variations. We solve this problem from the representation space perspective and propose Contrastive Decoupled Regularization (CoDeR), a module-level regularization method that guides representation learning using class prototypes as anchors without introducing any learnable parameters, steering hierarchical representations toward more discriminative directions. Specifically, for a target module $B_{i}$ , we maintain an independent cache that collects the module’s outputs and corresponding class labels during each training epoch. At the end of each epoch, features are aggregated by class and mapped onto a hypersphere to compute cluster centers as class prototypes. In the next epoch, these prototypes guide updates in module $B_{i}$ , pulling representations toward their ground-truth class prototypes and pushing them away from others. This strengthens inter-class separation in the representation space and directly addresses the core challenge of fine-grained recognition. Furthermore, we apply CoDeR in parallel across multiple modules to accelerate information propagation to earlier layers. This enables shallow layers to learn semantically meaningful representations earlier in training and mitigates the delayed representation update problem. Overall, CoDeR provides a simple, general, and effective supervised regularization mechanism that demonstrates the value of imposing constraints on high-dimensional representations. We conduct extensive experiments on ImageNet, six fundus medical imaging datasets, and two standard semi-supervised learning benchmarks. Consistent improvements across all settings validate the effectiveness and cross-domain applicability of our method.
Förderung
- National Natural Science Foundation of China NSFC
- Noncommunicable Chronic Diseases–National Science and Technology Major Project
- National Key Research and Development Program of China
- Shanghai Municipal Science and Technology Major Project
Zitieren
APA
Wang, Z., Li, T., Qin, Y., & Sheng, B. (2026). Contrastive Decoupling: Dynamic Regularization for Enhanced Fine-Grained Image Classification. IEEE Transactions on Circuits and Systems for Video Technology, 36(6), 7588-7599. https://doi.org/10.1109/tcsvt.2026.3661760
Vancouver
Wang Z, Li T, Qin Y, Sheng B. Contrastive Decoupling: Dynamic Regularization for Enhanced Fine-Grained Image Classification. IEEE Transactions on Circuits and Systems for Video Technology. 2026;36(6):7588-7599. doi:10.1109/tcsvt.2026.3661760
BibTeX
@article{Wang2026,
author = {Wang, Zheyuan and Li, Tingyao and Qin, Yiming and Sheng, Bin},
title = {Contrastive Decoupling: Dynamic Regularization for Enhanced Fine-Grained Image Classification},
journal = {IEEE Transactions on Circuits and Systems for Video Technology},
year = {2026},
volume = {36},
number = {6},
pages = {7588--7599},
doi = {10.1109/tcsvt.2026.3661760}
}
EUVIMED