| Squeezed light and entangled light are very important resources in quantum optics,and have attracted much attention from researchers due to their application prospects in many fields.The initial research on these two kinds of light focused on the fundamental transverse mode Gaussian beam,but with the development of the field of quantum information processing,the fundamental mode can no longer meet the demand.Compared with the fundamental mode,the spatial distribution of the higher-order mode is more complex and has a higher degree of freedom,so the higher-order mode has greater advantages,so scholars began to turn their attention to the preparation of the higher-order mode.Higher-order modes can be prepared by a beam shaping system based on a spatial light modulator,and the key is to generate a hologram loaded on the spatial light modulator.Among the various methods of generating holograms,computer-generated hologram technology has attracted much attention,and the optimization algorithm of computer-generated holograms is a research hotspot today.In recent years,with the rise of deep learning technology,more and more researchers have begun to apply deep learning to computer-generated holograms to optimize algorithms for computer-generated holograms.However,most models currently applied to computer-generated holograms are based on convolutional neural networks,and other excellent deep learning models are worth trying.Based on the preceding statement,this paper relies on the actual optical shaping system to study the deep learning computer generated hologram scheme based on the Vision Transformer network model.Specifically include:(1)Integrating the optical shaping system with computer-generated hologram technology,and using deep learning technology for optimization,a computer-generated hologram model based on Vision Transformer is proposed.In this model,the physical process of the beam shaping system is innovatively added,so that the computer-generated hologram is combined with the hardware.(2)The proposed model is trained using the Mnist handwritten digit dataset,and then the trained model is used to predict the test set to generate a hologram,and then the hologram is used to reconstruct the target light field.The reconstructed images verify the feasibility of the model in generating holograms to realize the target light field reconstruction problem.The quality of the reconstructed image is evaluated using the SSIM index,and the results show that the average SSIM value of the reconstructed image of the hologram corresponding to all the images in the test set is 0.751.This paper enriches research on the application of deep learning in the field of computer-generated holograms,provides a new solution for the preparation of high-order modes,and offers a new thinking direction for generating higher-quality holograms in the future. |
