Autocorrelation-guided dual-modality fusion network for high-fidelity and robust scattering imaging

Learning-based scattering imaging usually relies on either speckle patterns or autocorrelation maps, leading to a trade-off between fidelity and robustness. Here, we propose an autocorrelation-guided dual-modality framework that combines stable structural priors with speckle-based details. A physics-consistent autocorrelation crop augmentation (ACA) strategy is introduced to improve cross-domain robustness. Cross-modal fusion and two-stage progressive training further enhance reconstruction quality. Experiments show that our method achieves more accurate and robust scattering imaging than existing approaches.