Low-light lensless image enhancement via diffusion model

This paper introduces a deep learning-based low-light lensless image reconstruction enhancement algorithm that can effectively reconstruct low-light lensless images and achieve significant enhancement effects under low-light conditions. The algorithm comprises two stages: (1) A preliminary imaging stage, where the forward imaging model is utilized as a prior knowledge to obtain the initial reconstruction results; (2) A perceptual enhancement network built upon the conditional diffusion model for detailed enhancement, resulting in realistic images with normal lighting and reduced noise. Experimental results on simulated datasets demonstrate that this algorithm exhibits superior reconstruction performance with Learned Perceptual Image Patch Similarity (LPIPS) and Structural Similarity Index Measure (SSIM) up to 0.0887 and 0.7454, respectively, and successfully realizes the reconstruction and enhancement of lensless low-light images.