[End-to-end digital holographic phase reconstruction using an improved convolutional neural network]

Phase reconstruction is a critical step in digital holographic microscopy, aiming to recover the phase distribution of a specimen from its recorded holographic interferogram. To address the limitations of existing methods in preserving weak phase boundaries and reconstructing local structures, this study proposes a structure-enhanced convolutional neural network termed AG-Net. The network employs multiscale convolutions and Ghost residual blocks to improve feature extraction efficiency. CombinedAttention modules are introduced into the skip connections and decoder to select and enhance informative features, while SubPixel convolution is used to improve the reconstruction of phase boundaries and local details. On the MNIST-based simulated dataset, AG-Net achieved an structural similarity index of 0.864, a peak signal-to-noise ratio of 26.43 dB, and a Pearson correlation coefficient of 0.972. On the real red blood cell dataset, the corresponding values were 0.810, 18.73 dB, and 0.810, respectively, outperforming the selected comparison methods. The results demonstrate that AG-Net improves structural consistency, reconstruction accuracy, and boundary continuity, indicating its effectiveness for end-to-end digital holographic phase reconstruction.