Remote Sensing Image Dehazing Using Adaptive Region-Based Diffusion Models

Abstract

Remote sensing (RS) dehazing is a long-standing topic, as the contrast and clarity of images are seriously reduced in the severe environment. Recent learning-based ways achieve great success in RS dehazing, while these models usually suffer from unusual artifacts, mode collapse and size limitation. To address these problems, we propose an Adaptive Region-Based Diffusion Dehazing Net (ARDD-Net), which is a novel diffusion-based model toward free-form RS dehazing. To be specific, a guided diffusion process is employed to transform clear RS region parts by adding Gaussian noise repeatedly, and then restore the clear ones through a reverse generative procedure. In addition, Region-Based model is employed to address the free-size images for diffusion model. Further, a cyclic shift strategy is applied to the inference process, in order to avoid the inconsistent effect of certain regions and increase the image harmony. With extensive experiments, we state that our model gets outstanding performances than the comparing methods on both synthetic and real-world RS images.