Physics-enhanced deep learning for dual-branch single-pixel vision via highly scattering media

Single-pixel imaging (SPI) offers sensitivity for optical sensing, but dynamic scattering media severely degrade the captured 1D bucket signals, causing traditional image reconstruction algorithms to fail and hindering target recognition. To overcome the resulting sim-to-real gap, we propose a physics-enhanced dual-branch digital twin framework for single-pixel vision in highly scattering environments. By algorithmically embedding physical channel degradations-including optical blooming, spatial misalignment, and dynamic turbulence scintillation-into the synthetic training pipeline, the network learns to extract robust, noise-invariant features directly from heavily corrupted measurements. The architecture utilizes a shared encoder that simultaneously drives a semantic classification branch for rapid target recognition and a parallel spatial reconstruction branch for visual validation. Experimental validations in a dynamically turbid underwater testbed demonstrate that the proposed dual-branch model significantly outperforms conventional mathematical methods, e.g., differential ghost imaging and total variation, in structural fidelity and noise suppression. Furthermore, the framework exhibits exceptional compressive sensing efficiency, achieving a peak classification accuracy of 81.2% at a highly compressed sampling ratio of just 19.53%. The system maintains robust semantic recognition even under extreme physical turbulence, where conventionally reconstructed targets are rendered completely unrecognizable. This physics-enhanced deep learning approach provides a highly efficient solution for real-time optical sensing and target identification in dynamically challenging macroscopic channels.