.png){kind=icon}
.png){kind=icon}
Return
Diffusion Posterior Sampling for SAR Despeckling
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2025.3541013.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Despeckling is an important preprocessing step for synthetic aperture radar (SAR) image understanding and interpretation, and it can be regarded as an ill-posed inverse problem that heavily depends on the exploration of proper priors on clean images. Motivated by remarkable capability of deep generative priors, we aim to leverage diffusion denoising probabilistic model (DDPM) to achieve a good tradeoff between speckle suppression and detail preservation. To do so, we design a diffusion posterior sampling scheme based on the approximation of likelihood score function via estimation of reverse transition distribution and utilization of multiplicative noise model. The proposed method has two distinct superiorities and the first one lies in the pretrained model, escaping from the limitations of the pairs of clean and speckled images and the high computation complexity of supervised training. More importantly, the proposed method needs no transforms of SAR images, avoiding changing the statistical distribution of speckle and approximating the complex distribution in transform domain. Finally, we empirically validate the proposed method on synthetic and real SAR images, and the visual effects and the quantitative evaluation of the experimental results indicate the superiority of the proposed method over the state-of-the-art compared approaches. Besides, the robustness of the proposed method is also validated via analyzing its performance under the conditions of different speckle intensities, different initial samples (ISs) of diffusion models, and different values of the algorithm parameters.
Keywords:
Speckle
Radar polarimetry
Noise
Training
Transforms
Optical filters
Image edge detection
Noise reduction
Inverse problems
Despeckling
diffusion model
likelihood score function
posterior sampling
pretrained model
pretrained model
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W