Structure-boosted de-scattering reconstruction framework for high-fidelity speckle-correlation optical scattering imaging

Abstract

Speckle-correlation optical scattering imaging is an emerging technology for imaging through turbid media without requiring prior knowledge of the medium's scattering properties. However, achieving high-fidelity reconstructions remains challenging due to noise corruption in captured speckle patterns originating from both ballistic light interference and diffuse scattering disturbances. While existing denoising methods have demonstrated success in improving reconstruction quality, limitations persist in ideal scattering envelope fitting, system complexity reduction, and the need for extensive training-validation datasets. Herein, a structure-boosted de-scattering reconstruction (SBDR) framework was proposed to address these challenges, which integrates the computational steps into the physical modeling and consists of three modules: unrolled phase retrieval, object structure extraction and generation, and diffusion-based detail recovery. Experimental validation under strong scattering conditions (PSNR < 2.1 dB) demonstrates the framework's robustness in achieving high-fidelity reconstructions and its transferable adaptability to diverse scattering imaging configurations. Furthermore, quantitative comparisons with state-of-the-art network-based de-scattering methods reveal that SBDR achieves an average improvement of more than 20% in reconstruction fidelity while reducing the required training dataset size by over 96%.