关键词: 遥感图像, 语义分割, 小样本学习, 原型学习

Abstract: Deep learning-driven semantic segmentation methods for remote sensing images rely heavily on a large number of manually labeled samples and exhibit poor generalization for unknown tasks, especially in the fine-grained semantic segmentation task where the category system is constantly updated, and the recognition accuracy of the unknown categories (the categories that don't exist in the training samples) needs to be urgently improved. Based on this, the paper proposes a cross-category few-shot segmentation method aimed at multiple farmland categories. The method designs a dual-branch structure, comprising a support branch and a query branch, where the support branch is used for the extraction of segmentation prior, and the query branch is used to complete the propagation of segmentation prior and obtain the segmentation results of the query image. Additionally, the method applies query features to generate self-supporting query prototypes, which significantly improves the expressive ability of the prototypes; a regularization mechanism for prototype alignment between the support and query set is introduced, which makes full use of the knowledge from the support set and improves the discriminative ability of the segmentation. The experiments simultaneously introduce high spatial resolution and hyperspectral image land cover datasets to fully validate the performance of the proposed method. The experimental results show that compared with the existing few-shot segmentation methods, the proposed method can obtain more excellent cross-category farmland recognition results under few-shot conditions.

Key words: remote sensing images, semantic segmentation, few-shot learning, prototype learning