Abstract

Few-shot learning, which aims to learn the concept of novel category from extremely limited labeled samples, has received intense interests in remote sensing image scene classification. Most of the existing methods inherit the philosophy of prototype learning and tackle classification as the prototype-based metric matching problem. Despite the achievement that has been obtained so far, the problems of interclass metric misalignment and intraclass variations have become two main challenges that obstacle the performance. In this paper, a novel transductive learning framework with conditional metric embedding is proposed to remedy these problems. First, a conditional metric embedding mechanism is introduced to perform anisotropic embedding for each pair of the support category and query instance. This design provides the model with flexible scalability to accommodate the metric biases across classes. Second, a transductive prototype learning strategy is presented to enhance the robustness of the prototype against intraclass variations. The unlabeled query instances are transformed into pseudo instances equipped with credibility coefficients and then leveraged to calibrate the prototype estimation bias in low data regimes. Third, a long-term consistency regularization is designed, which continuously memorizes the historical prototypes to generate additional supervision of interclass separation in the global label space. Benefiting from this design, the discriminability of the prototypes obtains obvious improvement. Finally, extensive experiments are conducted on three public benchmark remote sensing datasets. The experimental results demonstrate the validity and the superiority of the proposed method in low-shot conditions.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call