Phase retrieval network for multi-functional holography based on self-supervised learning

Phase-only hologram reconstructs desired field distribution by modulating the phase of incident light, which is an important method for realizing full-color or 3-dimensional (3D) display. Under a given condition, traditional iterative algorithms or data-driven neural networks are employed to retrieve the phase profiles of full-color or 3D holograms. However, traditional iterative algorithms are required to repeat the optimization process for different target images, while the performance of data-driven neural network is limited by the training dataset. In this work, we propose a phase retrieval network for multi-functional holography based on self-supervised learning. The proposed method based on physics inspired network is formed by image preprocessing, deep neural network (DNN), and diffraction propagation. After self-supervised training, the proposed network retrieves the phase profile by inputting the target images, working wavelengths, and image distances. Compared with traditional iterative algorithms, the proposed network accelerates the phase retrieval process, suppresses the sparkle noise and improves the quality of reconstructed images. This proposed method has a great potential to realize real time 3D full-color display.