Few-shot image classification (FSL) aims to recognize novel classes with extremely limited labeled samples, which poses a critical challenge to traditional deep learning methods relying on massive annotated data. Meta-learning and metric learning have emerged as promising paradigms to address this issue, but existing approaches still suffer from prototype instability, insufficient feature discriminability, and poor generalization to domain shifts. To overcome these limitations, this paper proposes a Meta-Learning Enhanced Hybrid Framework (MLE-HF) that integrates prototypical networks (PN) with contrastive learning (CL) for few-shot image classification. Specifically, we design a dual-branch architecture: the primary branch leverages meta-learned prototype generation to capture task-specific semantic representations, while the auxiliary branch employs supervised contrastive learning to enhance feature separability in the embedding space. We introduce a dynamic prototype calibration mechanism based on expectation-maximization (EM) algorithm, which iteratively refines class prototypes using both support set samples and high-confidence query set pseudo-samples. Moreover, a novel Meta-Contrastive Loss (MCL) is proposed to align meta-training and meta-testing distributions, ensuring the transferability of learned features. We benchmark MLE-HF on miniImageNet, tieredImageNet, CUB-200-2011 and FC100 under 1-shot and 5-shot protocols. It sets a new state-of-the-art, reaching 58.71 % on 5-way 1-shot and 76.39% on 5-way 5-shot miniImageNet. Ablations pinpoint the contribution of every module, while visualizations reveal sharper, more robust embeddings. By intertwining metric and contrastive losses inside a meta-learning loop, the framework offers a fresh recipe for accurate few-shot recognition in data-scarce settings.
Xinwei Liu (Sun,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: