Research On Computational Holographic Network Based On Score Matching Generative Model

Computational holography technology is a technology that uses a computer to simulate the wavefront recording process in traditional optical holograms,and realizes holographic display by loading computational holograms onto spatial light modulation devices.In recent years,computational holography technology has developed rapidly and has been successfully applied in three-dimensional display,holographic measurement,optical storage,etc.,but the computational holographic algorithm still has problems such as poor quality of generating holograms.With the major breakthroughs made by deep learning in the fields of computer vision and image processing,its application to computational holography has gradually become a research hotspot.However,the current deep learning methods used in computational holography generally have some problems,such as the model is not stable enough to train and the robustness is insufficient.In addition,most deep learning methods face problems such as lack of data consistency and lack of datasets when generating holograms.In order to solve the above problems,this topic designs a computational holographic network based on the score matching generative model,which effectively improves the quality of the generated hologram.The main works are as follows:(1)In this dissertation,a computational holographic network based on score matching generation model is proposed to solve the problems of insufficient stability and difficulty in training and insufficient robustness.Firstly,the network constructs a scoring function to estimate the gradient of the probability density distribution of the data with high accuracy,and secondly,uses the efficiency of the numerical solver to sample the data from the conditional distribution of the given observation data,so that the method only needs to use the learned prior information to generate a high-quality pure phase hologram.Experimental results show that the algorithm achieves more accurate generation performance,and the imaging quality is higher than that of traditional generative holographic algorithms.(2)Aiming at the problem of insufficient network training caused by the lack of pure phase hologram dataset of computational generation holographic algorithm,this dissertation preprocesses 78227 and 90569 standard amplitude maps in LSUN-bedroom dataset and LSUN-church dataset respectively to generate pure phase holograms required for network training.By preprocessing these data,it is ensured that the generated holographic algorithm has sufficient training data,which improves the network training effect and provides important support for further improving the computational generation holographic algorithm.(3)Aiming at the problem of lack of data consistency items in the process of model generation of holograms,this dissertation combines the prior information learned by the model with the phase information generated by amplitude complex decomposition by combining subunits,so that the two phases are encoded into a pure phase hologram,so as to effectively avoid the problem of lack of data consistency terms in the experiment.In addition,in order to verify the effectiveness of the proposed network,an optical holographic imaging system is built to experiment with the pure phase hologram generated by the computational holographic network and the traditional algorithm.The actual optical debugging results show that the computational holographic network realizes the generation of high-quality holograms,which has certain practical application value.