Feature Reconstruction-guided Transductive Few-Shot Learning with Distribution Statistics Optimization

Abstract

The Transductive Few-Shot Learning (TFSL) method significantly enhances the recognition performance of few-shot learning models by leveraging the statistical data from query set samples. However, existing TFSL methods typically rely on global sample embeddings, overlooking class-level knowledge representations and spatial details. To address this, we propose a Feature Reconstruction-guided transductive few-shot learning method with Distribution Statistics Optimization (FR-DSO). Specifically, we have designed an Iterative Prototype-based Feature Reconstruction Module (IPFRM), which reconstructs query sample features using support set features and iteratively refined prototype features. Reconstruction errors across different classes serve as distance measures for classifying unlabeled samples. During fine-tuning, we utilize IPFRM to output high-quality pseudo-labels to achieve a stable optimization of the distribution of support set class features. Extensive experiments on mini-ImageNet, tiered-ImageNet, CUB-200-2011, and Aircraft benchmarks demonstrate the superior classification performance of our approach.