Denoised generative fusion networks for noise-robust few-shot image classification

Few-shot image classification (FSC) must often contend with inaccurate annotations and biased representations that arise when only a handful of training samples are available. These imperfections severely erode the reliability of existing FSC models. We propose Denoised Generative Fusion Networks (DGFN), a unified framework that tackles the twin challenges of label noise and representation bias in few-shot image classification. DGFN begins by pruning the support set with a principled uniqueness-aware composite distance that jointly considers intra-class coherence and inter-class separability, thus eliminating mislabeled or outlying samples without external supervision. A conditional variational auto-encoder is then trained on the cleaned support instances to hallucinate high-fidelity features, enriching class priors and effectively expanding the sample budget. Finally, an episodic fusion head adaptively weights real and synthetic features so that augmentation gains translate into reliable generalization. Extensive experiments on miniImageNet, tieredImageNet, CIFAR-FS, and CUB-FS show that DGFN boosts 5-way 1-shot accuracy by up to 4.6% over the state of the art and maintains its advantage under 20%$$\sim $$40% synthetic noise and cross-domain shifts, demonstrating its promise for safety-critical applications where clean, abundant annotations are unattainable.