Abstract:

With the rapid development of remote sensing satellites and unmanned aerial vehicle technologies， the acquisition of remote sensing images has become more convenient， making precise interpretation of image content a research hotspot. However， existing remote sensing image semantic segmentation methods struggle to handle new object categories that emerge due to scene changes when only a small number of samples are available， thereby affecting the overall performance of the model. To address the potential issues of new classes and new scenes in remote sensing scenarios， this paper introduces the generalized few-shot segmentation technique， aiming to rapidly adapt to new classes using a few samples and efficiently complete the semantic segmentation task for remote sensing images. Firstly， to tackle the challenge of distinguishing objects of different categories and different sizes in remote sensing images， a multi-scale feature fusion decoder is introduced to enhance the model’s segmentation ability for targets of different scales. Then， to address the background semantic shift problem in the generalized few-shot segmentation task， a background consistency modeling training strategy is designed to ensure the accurate expression of background features by the model. Furthermore， this paper proposes an inter-class relationship mining technique， which utilizes prototypes generated from a large amount of base class data to enhance the prototypes generated from a small amount of new class data， enabling them to acquire beneficial information from the base classes and achieve more precise segmentation. Finally， comparative experiments with mainstream methods are conducted on public datasets， and the mIoU segmentation metric is improved by up to 11.48%， demonstrating the effectiveness of the proposed algorithm.

Key words: remote sensing, semantic segmentation, generalized few-shot, knowledge distillation, relationship mining