| Abstract: |
| Clouds are ubiquitous in remote sensing images. Most of the existing methods for cloud removal are limited to either implementing on multi spectral images or exploiting supervised learning technique. In this paper, we propose an unsupervised diffusion approach by deploying the null space learning. The proposed approach is built upon two trained denoising diffusion probabilistic models by diverse remote sensing datasets so as to tackle the mixture of data from different sources. The simplified degradation and self-adaptive generalized inverse matrices are devised for the null space decomposition. For the diffusion model with null space decomposition, we derive its continuous reverse-time stochastic differential equation (SDE), which is theoretically proven to be variance preserving. We further derive the explicit formula for the expectation of the reverse-time SDE, which is conducive to algorithm improvement. As a byproduct, the proposed approach can also be applicable to the transparency separation. Numerical experiments on some remote sensing images demonstrate that the proposed approach outperforms some state-of-the-art unsupervised, even supervised, cloud removal methods. |
|