StableCam: Lensless imaging with stable diffusion model-based reconstruction

Abstract

Lensless imaging offers low-cost camera miniaturization by replacing traditional lenses with coded masks, enabling applications across diverse domains. However, its reconstruction quality is often constrained by the ill-conditioned nature of the inverse problem. In this paper, we present StableCam, a lensless imaging system employing a Modified Uniform Redundant Array (MURA) mask and a physically interpretable reconstruction network, StaRNet. StaRNet integrates a separable initialization module based on the lensless imaging model, which provides preliminary reconstructions with reduced computational complexity. Additionally, it includes a parameter-efficient image enhancement module that improves resolution and refines details. The experimental results demonstrate that the proposed StableCam enables to achieve high-quality reconstruction under both calibration and random matrix conditions. Compared to the previous lensless method, e.g., FlatNet, our approach achieves improvements across all evaluation metrics, including peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), perceptual image patch similarity (LPIPS), and Fréchet inception distance (FID).