Combining ResNet and Diffusion Model for ISAR Image Super-Resolution

Recently, denoising diffusion probabilistic models (DDPMs) have emerged as powerful tools for image generation. In inverse synthetic aperture radar (ISAR) super-resolution (SR) reconstruction, existing generative models like GANs often suffer from unstable training and mode collapse, whileas other deep learning methods often fail to recover weak scattering points and introduce artifacts. To tackle these challenges, we propose RDSRN, a novel ISAR super-resolution network combining ResNet and DDPM. RDSRN not only overcomes the limitations of generative networks but also enhances the ability to reconstruct scattering points. Initially, ResNet is employed to suppress sidelobe interference and preliminarily restore the low-frequency components of the image. The DDPM then focuses on recovering high-frequency details to generate more refined scattering point features. Specifically, the ResNet incorporates an attention mechanism and is trained using a combination of L1 loss and gradient consistency loss. The design above ensures that the recovered low-frequency components align with the physical properties specific to ISAR. Besides, we design a learnable frequency domain filter (LFM) and a frequency domain guided cross attention module (FDGCA) for the DDPM to optimize high-frequency reconstruction. Compared to state-of-the-art methods, RDSRN demonstrates a significant advantage in the precision of recovering weak scattering points.