HoloMamba: Physics-Aligned State-Space Network for High-Fidelity Single-Shot Holographic Reconstruction

Single-shot digital holographic reconstruction is intrinsically ill-posed due to missing phase information, resulting in conjugate artifacts and degraded high-frequency fidelity, while existing convolutional and Transformer-based models inadequately capture long-range diffraction behavior and essential physical constraints. To address this challenge, we propose HoloMamba, a physics-aligned state-space framework for high-fidelity single-shot hologram reconstruction. The proposed model adopts a dual-branch hourglass architecture with bidirectional Mamba propagation, enabling efficient modeling of forward and backward spatial dependencies while preserving complex-domain consistency in linear time. To further incorporate physical priors, a spectral- and numerical-aperture-aware fusion strategy is introduced to adaptively balance global contextual modeling and directional state propagation across frequency bands. In addition, a curriculum-guided spectral regularization scheme is designed to progressively enforce band-wise consistency, thereby stabilizing high-frequency reconstruction during training. Extensive experiments on biological specimens and standard resolution targets demonstrate that HoloMamba consistently achieves superior amplitude and phase reconstruction accuracy compared with state-of-the-art convolutional and Transformer-based methods, while exhibiting enhanced robustness to axial misplacement and distribution shifts. These results validate the effectiveness of integrating structured state-space modeling with physics-informed spectral learning, and establish a principled approach for physics-aware holographic reconstruction in lensless microscopy and related computational imaging applications.