SAR Despeckling using a Denoising Diffusion Probabilistic Model

Malsha V. Perera, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, Vishal M. PatelarXiv 2022 · DOI · OpenReview

TL;DR

This paper introduces SAR-DDPM, the first application of Denoising Diffusion Probabilistic Models (DDPMs) for Synthetic Aperture Radar (SAR) image despeckling. By leveraging a novel conditional noise predictor and a cycle spinning inference strategy, SAR-DDPM significantly outperforms existing methods in removing speckle noise while preserving crucial image details.

What Problem Does This Paper Address?

Synthetic Aperture Radar (SAR) is a powerful imaging technology, but its images are plagued by "speckle" – a signal-dependent, multiplicative noise. This noise significantly degrades image quality, making it harder to interpret the data for applications like disaster monitoring, land classification, and automatic target recognition. While many methods exist to "despeckle" SAR images, they often struggle to strike a balance between noise reduction and preserving fine details, leading to overly smoothed images or residual noise. This paper aims to address this challenge by proposing a new, more effective approach to SAR despeckling.

Methods

The core of the proposed method, SAR-DDPM, is based on Denoising Diffusion Probabilistic Models (DDPMs). DDPMs work by defining a two-stage process:

  1. Forward Diffusion: A clean image is gradually corrupted by adding Gaussian noise over a series of steps (T steps), eventually transforming it into pure noise.
  2. Reverse Diffusion: A neural network is trained to reverse this process, iteratively removing noise to recover the original clean image.

In SAR-DDPM, the key innovation is a conditional noise predictor. This predictor is trained to estimate the noise added at each step of the forward process, but crucially, it's conditioned on the speckled SAR image. This allows the model to learn the specific characteristics of SAR speckle. The network architecture is a U-Net, incorporating convolutional blocks, residual blocks, and attention mechanisms, with timestep embeddings guiding the denoising process.

To further enhance performance, the authors introduce a cycle spinning inference strategy. This technique involves cyclically shifting the input speckled image by various amounts, applying the trained DDPM to each shifted version, and then averaging the resulting despeckled images. This ensembling approach helps to mitigate distortions and improve the preservation of fine structural details.

Datasets

  • Training: The SAR-DDPM model was trained on 15,000 remote sensing images from the publicly available DSIFN dataset. These images were synthetically speckled using the multiplicative speckle model described in the paper (Eq. 1 and 2) with a single-look assumption (L=1).
  • Evaluation: The model was evaluated on both synthetically speckled images (from the DSIFN dataset) and real SAR images. For real SAR images, the paper uses Sentinel-I images.

Key Insights and Ideas

  • DDPMs for SAR: This paper demonstrates that DDPMs, which have shown great success in image synthesis and restoration, are highly effective for the challenging task of SAR despeckling.
  • Conditional Denoising: Conditioning the noise predictor on the speckled image is crucial for learning and removing SAR-specific noise.
  • Cycle Spinning Synergy: The cycle spinning technique, when combined with DDPMs, significantly boosts performance by averaging out potential artifacts and improving detail preservation. This suggests that ensembling diverse but related estimates can be a powerful strategy.
  • Balancing Act: SAR-DDPM appears to achieve a better balance between noise reduction and detail preservation compared to many existing methods, which often sacrifice one for the other.

Key Results

  • Quantitative Improvements: SAR-DDPM achieved state-of-the-art performance on both synthetic and real SAR images.
  • On synthetic data, it significantly outperformed methods like PPB, SAR-BM3D, ID-CNN, Trans-SAR, SAR-ON, and SAR-CAM in terms of average Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For example, SAR-DDPM with cycle spinning achieved a PSNR of 29.42 dB and an SSIM of 0.81, outperforming the next best method (SAR-CAM) by a notable margin.
  • On real SAR images, SAR-DDPM with cycle spinning yielded the highest Equivalent Number of Looks (ENL) values across multiple regions, indicating superior noise reduction while preserving image homogeneity.
  • Qualitative Improvements: Visual results show that SAR-DDPM effectively removes speckle noise while preserving fine edges, textures, and structural details that are often lost by other methods. The cycle spinning strategy further refines these results, reducing distortions and enhancing detail recovery.

Limitations and Future Work

  • Computational Cost: While DDPMs offer excellent results, they are known to be computationally intensive, especially during inference due to the iterative denoising process. The paper mentions training on an NVIDIA RTX 2080Ti GPU for 30k iterations.
  • Distortions in Homogeneous Areas: The paper notes that SAR-DDPM, while good at detail preservation, can sometimes introduce slight distortions in homogeneous areas, leading to lower ENL values compared to some smoothing-focused methods before cycle spinning. Cycle spinning helps to mitigate this.
  • Future Work: The authors suggest that their work opens avenues for further research into applying diffusion models for SAR image processing tasks. Potential future directions could involve exploring more efficient inference strategies for DDPMs or investigating their use in other SAR image understanding applications.

Summary updated

Abstract

Speckle is a multiplicative noise which affects all coherent imaging modalities including Synthetic Aperture Radar (SAR) images. The presence of speckle degrades the image quality and adversely affects the performance of SAR image understanding applications such as automatic target recognition and change detection. Thus, SAR despeckling is an important problem in remote sensing. In this paper, we introduce SAR-DDPM, a denoising diffusion probabilistic model for SAR despeckling. The proposed method comprises of a Markov chain that transforms clean images to white Gaussian noise by repeatedly adding random noise. The despeckled image is recovered by a reverse process which iteratively predicts the added noise using a noise predictor which is conditioned on the speckled image. In addition, we propose a new inference strategy based on cycle spinning to improve the despeckling performance. Our experiments on both synthetic and real SAR images demonstrate that the proposed method achieves significant improvements in both quantitative and qualitative results over the state-of-the-art despeckling methods.
BibTeX
@misc{perera2022sar,
  author = {Malsha V. Perera and Nithin Gopalakrishnan Nair and Wele Gedara Chaminda Bandara and Vishal M. Patel},
  title = {{SAR Despeckling using a Denoising Diffusion Probabilistic Model}},
  year = {2022},
  eprint = {2206.04514},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2206.04514},
}
Figure