Feature Transductive Distribution Optimization for Few-Shot Image Classification

Few-shot learning (FSL) requires vision models to quickly adapt to brand-new classification tasks with changing task distributions in the presence of limited annotated samples. However, the learned model is susceptible to overfitting and may fail to identify effective classification boundaries due to the biased distribution resulting from a limited number of training samples. Moreover, if the support samples from different classes in the new task are in close proximity, this may lead to fuzzy or even biased class decision boundaries. To address the issues, we propose a generation-based Feature Transductive Distribution Optimization (FTDO) in our research. Specifically, we calibrate the distribution of novel classes by utilizing high-confidence unlabeled query samples from these novel classes, together with the statistics of similar base classes, to generate a sufficient number of virtual training samples. In addition, we introduce a task commonality removal and discriminability enhancement module, which eliminates commonality from all features in the task along the task-commonality direction, and reinforces the retained discriminative features through a channel transformation function. Our method can be implemented using off-the-shelf pre-trained feature extractors and classification models, without requiring additional parameters. Experiments conducted on four few-shot classification datasets substantiate the superiority of our proposed method.