A conditioned feature reconstruction network for few-shot classification

Few-shot classification is one of the most daunting challenges in deep learning. The complexities of this task arise from the fact that category targets are often embedded within intricate and diverse background pixels, resulting in inconspicuous category features. Moreover, obtaining common category characteristics from a limited number of samples is difficult. Compounding the issue, models encounters categories that they have never seen before, rendering the prior guarantee of interclass variance infeasible. To address these dilemmas, this paper leverages the apriori conditioned information of few-shot tasks and introduces a Conditioned Feature Reconstruction Network (CFRN). The CFRN employs prototype reconstruction to minimize the prototype similarity among different classes and query reconstruction to maximize the similarity of (query, prototype) feature pairs. This approach increases the interclass variance while decreasing the intraclass variance, thereby enhancing separability and improving the saliency of the target features. An experimental validation demonstrates the effectiveness of the CFRN, which obtains state-of-the-art results on the mini-ImageNet, tiered-ImageNet, and CUB datasets.