Research On The Generation Of Computer-generated Holograms Based On Deep Learning

The hologram obtained by computer simulation of the interference and diffraction processes of light waves is called a computational hologram,which not only allows complete recording and reproduction of the amplitude and phase of light waves,but also has the advantage of low noise and high reproducibility.Compared with conventional optical holography,it can also generate holograms of virtual objects.Due to the high optical efficiency of phase modulation,pure phase-based holograms are a better choice for holographic displays in most cases.However,traditional iterative optimization-based phase hologram generation algorithms have an inherent trade-off between computational speed and accuracy,which limits the application of computational holograms for real-time displays.In recent years,deep learning has gradually emerged as an enabling tool for computational holography and other optical imaging applications,offering new options in the problem of fast processing of optical information.This paper develops an algorithmic and experimental study of computational holography based on deep learning.The main contents of the work in this paper are specified as follows.(1)A convolutional neural network model with the addition of residual blocks is constructed for the generation of pure phase computation holograms,which increases the recognition of features by the network and improves the training ability of the network and reduces the cost of computation time by sampling feature capture of the input image and preventing the network from overfitting after nine residual blocks.And the loss function that can effectively represent the image features is designed for the training content,which makes the network converge after iteration and completes the training of the network.The time to generate the hologram for the completed training network is only 0.1889 s,which is one order of magnitude faster than the traditional iterative angular spectrum method algorithm in terms of generation speed,which improved the impact of inherent trade-off issues.(2)The corresponding network input modules are constructed for 2D,3D and color images,and the datasets for the above three types of images are designed and produced.A dynamic weighting adjustment factor is designed to be introduced in the iterative angular spectrum method to calculate the label data(ground-truth holograms),which enables the reproduced image quality to be adjusted and optimized in the iterative process.By using the idea of intermediate values,the complex holographic surface complex amplitude is converted into two output parts with clearer correspondence and lower complexity,which makes the network easier to train the mapping relationship between input and output,and the training results of the real and imaginary parts of the complex amplitude in holographic plane can finally be mathematically calculated to generate phase holograms.The input data(target images)in the dataset is designed as four types including sparse dots,dense dots,large diameter circles and random size rectangular blocks in a total of 10,000 groups,which provides data guarantee for the generalizability of the network.(3)Numerical reproduction and photoelectric reproduction of the hologram imaging quality analysis were performed.The evaluation analysis was performed by root mean square error,peak signal-to-noise ratio,correlation coefficient and other evaluation indexes to verify the effectiveness of this research method.The photoelectric hologram reproduction experimental system was constructed with a pure phase-type liquid crystal spatial light modulator.Photoelectric reproduction experiments were conducted for holograms of two-dimensional,three-dimensional and color images,and the results of photoelectric reproduction experiments were obtained to experimentally verify the research method of this study.