NeOTF: Speckle-Constrained OTF Neural Representation for Non-invasive Imaging through Scattering Medium
Abstract
Imaging through scattering media via speckle correlation is fundamentally challenged by ill-posed reconstruction. To overcome this, we bypass direct object recovery and instead target the system's deterministic optical transfer function (OTF). We introduce NeOTF, a framework that learns an implicit neural representation of the OTF. By optimizing this representation using multi-frame speckle intensities and a physical Fourier-domain prior, NeOTF robustly retrieves the system's OTF. Subsequent deconvolution with this retrieved OTF yields high-fidelity object reconstructions. Both simulations and experiments demonstrate that NeOTF achieves superior accuracy and efficiency over conventional methods, establishing it as a practical solution for real-time scattering imaging.