Mean-Reverting Diffusion Model-Enhanced Scattering Imaging

Abstract

Scattering media disrupt the rectilinear propagation of light, significantly degrading the resolution and clarity of optical imaging systems. Current scattering imaging techniques usually focus on simple targets and present limitations in imaging quality and reconstruction efficiency. To address these limitations, a mean-reverting diffusion model-enhanced scattering imaging (MRDS) is proposed. During training, prior information is extracted by diffusing the training data into an intermediate state with stable Gaussian noise. Reconstruction begins with low-quality images from physically-guided inversion, followed by iterative solving of reverse-time stochastic differential equations via the Euler-Maruyama method, integrating learned prior information to efficiently reconstruct high-quality images. Simulative and experimental validations demonstrate that MRDS outperforms traditional methods in reconstructing images with fewer artifacts and enhanced detail clarity. Quantitative metrics further demonstrates excellent reconstruction performance, with average metrics reaching 41.19 dB for PSNR, 0.99 for SSIM and 0.0085 for LPIPS. The reconstruction time per image is 2.19 seconds, representing a 44.2-fold acceleration compared to conventional methods. The proposed method achieves high-quality reconstructions of complex targets in a significantly shorter time, which dramatically boosts the efficiency of scattering imaging.