SAR Image Synthesis with Diffusion Models

TL;DR

This work adapts denoising diffusion probabilistic models (DDPM) to synthetic aperture radar (SAR) data, addressing the scarcity of labeled SAR data by generating high-fidelity SAR images. It presents the forward diffusion process that adds Gaussian noise and a learnable reverse denoising process, trained to predict the noise $\boldsymbol\epsilon_\theta$, with a UNet backbone and optional class conditioning. The authors demonstrate that DDPMs outperform state-of-the-art GANs on the MSTAR dataset across multiple metrics (IS, FID, KID), and that pretraining on a large clutter dataset further improves image quality. This approach offers a practical path to augment SAR datasets for downstream tasks such as target recognition, with potential extensions to other SAR imaging tasks like translation and super-resolution.

Abstract

In recent years, diffusion models (DMs) have become a popular method for generating synthetic data. By achieving samples of higher quality, they quickly became superior to generative adversarial networks (GANs) and the current state-of-the-art method in generative modeling. However, their potential has not yet been exploited in radar, where the lack of available training data is a long-standing problem. In this work, a specific type of DMs, namely denoising diffusion probabilistic model (DDPM) is adapted to the SAR domain. We investigate the network choice and specific diffusion parameters for conditional and unconditional SAR image generation. In our experiments, we show that DDPM qualitatively and quantitatively outperforms state-of-the-art GAN-based methods for SAR image generation. Finally, we show that DDPM profits from pretraining on largescale clutter data, generating SAR images of even higher quality.

Figures (7)

• Figure 1: Training and inference phases for DDPM. During training, input SAR data infused with noise, timestep and class embeddings are fed to the UNet model, which learns to predict the noise added to the input samples. During inference, noise is sampled from a normal distribution, fed to the UNet and iteratively denoised to finally obtain a SAR image.
• Figure 2: DDPM processing illustrated with the MSTAR dataset. The model is comprised of two processes: a forward diffusion process that adds noise to the data and learns it for each timestep $t$; and a reverse denoising process that draws samples from a Gaussian noise at time $T$ and iteratively denoises its input to generate new data samples.
• Figure 3: The behaviour of $\bar{\alpha}_t$ for different noise schedulers.
• Figure 4: The clutter data depicts the surroundings of MSTAR targets, and illustrates scenarios such as fields, trees and roads.
• Figure 5: Images generated by class-conditioned generative networks. The images generated by cDDPM represent the background in a more accurate way and reconstruct the targets more precisely.