Physics-guided multi-scale deep learning network for incoherent digital holographic reconstruction

We propose a novel holographic reconstruction network termed HPD-Net that integrates physical priors and data-driven learning. Based on a multi-scale parallel encoder-decoder architecture, the network generates multi-scale physical prior intensity maps via a non-linear reconstruction module, achieves adaptive fusion of prior information and deep features through a dual-attention prior fusion (DAPF) module, and introduces learnable physical residual connections to reduce the learning difficulty of the network, thereby realizing end-to-end high-fidelity reconstruction of holographic images. The experimental results show that compared with the traditional nonlinear reconstruction method, the PSNR, SSIM and ECC of the proposed method are significantly improved on different datasets, and it is better than the reconstruction performance of the traditional deep learning method. This effectively enhances the reconstruction quality and robustness of incoherent holographic imaging.