Study On Near-Eye Hologram Calculation Algorithm Based On Deep Learning

Holographic display technology has been developing rapidly in recent years.This technology can be used to enable people to obtain a visual perception close to the real near eye scene.It has a very broad application prospect in military,education,medical and other aspects.As one of the most promising holographic display technologies,CGH technology has developed rapidly in recent years with the continuous iterative evolution of holographic optical theory and the continuous maturity of computer software and hardware technologies.Scientific research teams in various countries have made many breakthrough achievements in this field,at present,there are two research hotspots in this field:how to design optimization algorithms to better improve the image quality of holographic reconstruction;And how to speed up the generation of holograms,so as to effectively reduce the time cost in the process of generating holograms in large batches.This dissertation focuses on the above two hot issues to carry out relevant research and experimental verification.The main work is as follows:1.Several common holographic phase map generation algorithms suitable for near-eye display are studied.Firstly,this dissertation studies the common iterative algorithms such as Gerchberg-Saxton(GS),random gradient descent and Wirtinger Holography(WH),and the quality of the reconstructed image,such as peak signal-to-noise ratio(PSNR)and structural similarity,are analyzed.On the basis of these,an innovative method for image correction based on high-dimensional space is proposed.The experimental light path is set up to implement and test the prototype system of near-eye display.Experiments verify that the values of the above two indicators obtained by this method are significantly improved,and the quality of reconstructed image is also improved.2.To speed up the generation of holograms,a method combining the self-designed deep neural network model with traditional iterative algorithm is proposed to generate pure-phase holograms quickly,optimize the loss function in the neural network model,and generate pure-phase holograms by batch training of the neural network model.The quality and convergence rate of the reconstructed images are evaluated comprehensively.By validating on the experimental platform,this method significantly improves the speed of batch generation of holograms compared to the iterative phase-only hologram optimization algorithm.