ICSS: Semantic Segmentation of Remote Sensing Images Based on Image Inpainting and Contrast Self-Supervised

Abstract

Semantic segmentation of remotely-sensed images is a fundamental problem for remote sensing research. However, unlike natural images, remote sensing images cover a larger surface area, contain more complex object categories, while often have a similar visual features across different categories. These characteristics pose new challenges for semantic segmentation. To address the complexity of semantic segmentation in remote sensing images, we have proposed a self-supervised method called ICSS (Inpainting and Contrast Self-Supervised) to extract image information and improve the generalization ability of the segmentation model. Specifically, we utilize self-supervised method to train remote sensing images and obtain pre-trained weights that can adapt to the unique features of these images. These pre-trained weights are then applied to the semantic segmentation task. Our experiments on the Potsdam dataset demonstrate that the self-supervised method effectively enhances the accuracy of semantic segmentation.