Abstract
Synthetic Aperture Radar (SAR) images are significantly degraded by multiplicative speckle noise, making their analysis and interpretation challenging. Recently, Denoising Diffusion Probabilistic Models (DDPMs) have demonstrated success in image generation and image enhancement tasks, such as denoising and super-resolution. This paper examines the performance of DDPMs for SAR despeckling using both synthetically speckled and real SAR images. Proposed modifications to the DDPM framework include (i) using a non-uniform step size and spread, along with early stopping in the denoising process, and (ii) sample aggregation by training of a secondary aggregating U-Net to extract additional performance from the partially denoised DDPM samples. Both of the proposed modifications improve accuracy and reduce inference time by utilizing fewer iterations. Various datasets, training methodologies and evaluation metrics are utilized to comprehensively assess the effectiveness of DDPM models for SAR despeckling and benchmark their performance against state-of-the-art SAR despeckling techniques, focusing on accuracy, training time, and evaluation time. Our findings provide insights into the benefits and limitations of DDPMs in the context of SAR despeckling. While diffusion models for SAR produce sharper and more realistic imagery, they sometimes hallucinate and result in lower quantitative performance compared standard U-Net denoising, highlighting the need for better metrics and improved techniques.