Calibration-free 3D imaging through scattering media via diffusion-kernel-guided reconstruction

Three-dimensional (3D) imaging based on the diffusion equation typically requires prior knowledge of the medium's thickness and optical properties, thereby limiting its practical applicability. We present a calibration-free 3D imaging framework that overcomes this limitation. We formulate a semi-empirical forward model that captures both the backscattered signal from the medium and the target signal, and employ a multi-stage optimization strategy to estimate the medium parameters and decouple the target signal. Using these estimated parameters, we then construct a diffusion-kernel-guided spatially adaptive regularization to reconstruct the hidden target from the decoupled signal. This regularization suppresses noise while preserving structural features. Experiments across varying medium thicknesses and multi-target scenarios demonstrate that our method achieves robust reconstruction performance.