Self-Diffusion Driven Blind Imaging

TL;DR

DeblurSDI introduces a training-free blind imaging framework that uses a reverse self-diffusion process to jointly recover a sharp image and the PSF under optical aberrations and motion blur. It employs two untrained networks—a denoiser for the image and a kernel generator for the PSF—optimized over diffusion steps with a data-fidelity objective and a sparsity prior on the PSF. Empirical results on simulated aberrations and motion blur demonstrate state-of-the-art performance and robust kernel recovery, outperforming both training-free and pretrained prior-based baselines. The approach offers a calibration-free, robust solution for high-fidelity image restoration in practical optical systems, at the expense of longer runtimes due to iterative optimization.

Abstract

Optical imaging systems are inherently imperfect due to diffraction limits, lens manufacturing tolerances, assembly misalignment, and other physical constraints. In addition, unavoidable camera shake and object motion further introduce non-ideal degradations during acquisition. These aberrations and motion-induced variations are typically unknown, difficult to measure, and costly to model or calibrate in practice. Blind inverse problems offer a promising direction by jointly estimating both the latent image and the unknown degradation kernel. However, existing approaches often suffer from convergence instability, limited prior expressiveness, and sensitivity to hyperparameters. Inspired by recent advances in self-diffusion, we propose DeblurSDI, a zero-shot, self-supervised blind imaging framework that requires no pre-training. DeblurSDI formulates blind image recovery as an iterative reverse self-diffusion process that begins from pure noise and progressively refines both the sharp image and the blur kernel. Extensive experiments on combined optical aberrations and motion blur demonstrate that DeblurSDI consistently outperforms other methods by a substantial margin.

Figures (11)

• Figure 1: PSFs for individual and combined aberrations (defocus, coma, astigmatism, spherical, coma+spherical, defocus+coma). Uprow: wavefronts. Downrow: PSFs.
• Figure 2: Sensitivity of noise scheduled steps, $T$, and inner iterations, $S$. The graphs show the SSIM (left) and PSNR (right) scores for different numbers of outer diffusion steps ($T \in \{10, 20, 30, 40\}$) and inner optimization iterations ($S \in \{25, ..., 500\}$).
• Figure 3: Evolution of image and kernel estimates during DeblurSDI's reverse diffusion process.
• Figure 4: Ablaion study on the network for PSF estimation. This figure compares the performance of the "Standard" mode against the "Diffusion" mode architecture with a varying number of hidden layers ($n=1$ to $n=5$).
• Figure 5: Performance comparison under different addictive noise on the blurred image.