Phase reconstruction of inline holograms and three-dimensional topography visualization based on an all-optical residual diffraction neural network

Electronic computer-based neural networks can recover high-resolution quantitative phase images from inline holograms to reveal the fine structure of samples, which may face issues such as intensive computing, energy consumption, and time delay. In contrast, optical diffraction computing not only has super-fast parallel processing capabilities and extremely low energy consumption but also is naturally compatible with the physical process of holographic imaging. However, since existing optical diffraction computing models are not able to preserve the phase detail in phase reconstruction, they focus on intensity recovery of holograms. To address this problem, this paper proposes an all-optical residual diffraction neural network (RDNN) architecture to obtain a high-precision phase distribution of inline holograms. In this method, residual optical structures and cascaded diffraction layer designs are combined to effectively reduce the loss of phase information during reconstruction, thereby ensuring a high degree of consistency between the reconstructed phase and the original phase. A series of numerical simulation results shows that the proposed architecture can be generalized to the phase reconstruction task of three-dimensional biological samples. In phase reconstruction of human erythrocytes, the average SSIM and PSNR reach 0.82 and 28, respectively. Additionally, we assess the network's performance with noisy holograms, demonstrating that, unlike fully connected diffraction neural networks, this approach significantly enhances phase reconstruction quality, even under random Gaussian noise. These results indicate that the proposed RDNN has the potential to replace the electronic neural network to achieve accurate phase reconstruction of inline holograms and provide an efficient solution to observe the living cell dynamics.