On Denoising Diffusion Probabilistic Models for Synthetic Aperture Radar Despeckling

Area of Science:

• Remote Sensing
• Image Processing
• Artificial Intelligence

Background:

• Synthetic Aperture Radar (SAR) images suffer from speckle noise, hindering analysis.
• Denoising Diffusion Probabilistic Models (DDPMs) are emerging as powerful tools for image enhancement.

Purpose of the Study:

• To evaluate the effectiveness of DDPMs for SAR image despeckling.
• To propose and assess modifications to enhance DDPM performance for SAR data.

Main Methods:

• Utilized synthetically speckled and real SAR images for testing.
• Implemented DDPMs with modifications: non-uniform step size, early stopping, and secondary U-Net aggregation.
• Benchmarked against state-of-the-art despeckling techniques.

Main Results:

• Proposed modifications improved accuracy and reduced inference time.
• DDPMs produced sharper, more realistic SAR imagery.
• Quantitative performance sometimes lower than U-Net denoising due to hallucination.

Conclusions:

• DDPMs offer benefits for SAR despeckling, including improved visual quality.
• Limitations include potential hallucination and the need for refined evaluation metrics.
• Further research is needed to optimize DDPMs for SAR applications.