Citation

Lim, Jia Min and Lim, Kian Ming and Goh, Pey Yun and Lee, Chin Poo (2026) Diffusion-inspired prototypical network for few-shot fine-grained classification. Applied Intelligence, 56 (15). ISSN 0924-669X

Text s10489-026-07509-2.pdf - Published Version Restricted to Repository staff only Download (3MB)

Abstract

Distinguishing minute visual differences under conditions of extreme data scarcity remains a critical challenge in few-shot fine-grained image classification. When relying on limited support samples, existing methods frequently overfit to spurious background details rather than learning robust semantic representations, limiting their generalization ability. To address this issue, we propose Diffusion-inspired Prototypical Network (DiffProNet), a framework that incorporates a self-supervised feature-space denoising objective into the few-shot learning pipeline. Rather than performing image generation, the proposed method injects Gaussian noise into the support feature representations and employs a Feature Denoising Head to predict the injected noise through a self-supervised auxiliary objective. This denoising objective encourages the encoder to learn robust and discriminative feature representations, thereby regularizing the latent feature space and mitigating overfitting to irrelevant visual details. Extensive experiments on the CUB-200-2011, Stanford Dogs, and Stanford Cars benchmarks demonstrate the effectiveness of the proposed approach, yielding competitive performance under both 1-shot and 5-shot settings and validating the effectiveness of feature-space denoising for few-shot fine-grained image classification.

Item Type:	Article
Uncontrolled Keywords:	Few-Shot learning, Fine-Grained image classification, Self-Supervised learning, Prototypical network, Feature-Space denoising
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101-6720 Telecommunication. Including telegraphy, telephone, radio, radar, television
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	29 Oct 2026 06:54
Last Modified:	29 Oct 2026 06:54
URII:	http://shdl.mmu.edu.my/id/eprint/16992

Downloads

Edit (login required)