Diffusion Models Based Null-Space Learning for Remote Sensing Image Dehazing
Yufeng Huang, Zhiyu Lin, Shuai Xiong, Tongtong Sun
Abstract
Remote sensing (RS) dehazing is a challenge topic, as images captured under hazy scenarios often suffer from seriously quality degradation and inconsistency. RS image restoration has been significantly improved with the use of learning-based ways, while current methods are still struggling to restore the complex details for large irregular RS images with ununiform haze. In this letter, we propose an Adaptive Diffusion Null-space Dehazing Network named ADND-Net, which is a novel diffusion model based null-space learning toward free-form RS image dehazing. Specifically, a range-null space decomposition is applied to improve the reverse diffusion process for image consistence. With the help of range-null space content, we further advance the adaptive region-based diffusion module to address the unlimited-size RS images, and increase the dehazed image quality. Extensive experiments show that our designed model outperforms other comparing dehazing methods on both synthetic and real-world RS datasets.