Building Extraction from Remote Sensing Imagery Using Multi-feature Adaptive Sampling and Denoising Diffusion Implicit Model

The Denoising Diffusion Probabilistic Models (DDPM) have become a breakthrough technology in the field of computer vision. Besides excelling in generative modeling, they also demonstrate efficient performance in various application areas such as image restoration, super-resolution, editing, translation, synthesis and semantic segmentation, attracting widespread attention. In this paper, we explore a new method of using DDPM in the remote sensing field for building semantic segmentation. Our focus is on optimizing the sampling process of the segmentation target through a combination of multi-feature adaptive sampling strategy and the Denoising Diffusion Implicit Model (DDIM), thus achieving dynamic sampling. This method aims to improve the overall accuracy and edge precision of segmented images, while also enhancing sampling efficiency. Experimental results on the Vaihingen dataset show that our method, compared to the baseline model SegDiff, achieves an increase of 0.59% in mIoU and 0.85% in FBound, thereby optimizing the overall performance of the segmentation task.