卷积误差消除的模型驱动相位型全息图生成网络  

Convolution-error-free model-driven neural network for phase-only computer-generated hologram

作　　者：何泽浩 刘珂瑄 曹良才[2] 张岩[1] HE Zehao;LIU Kexuan;CAO Liangcai;ZHANG Yan(Department of Physics,Capital Normal University,Beijing 100048,China;Department of Precision Instrument,Tsinghua University,Beijing 100084,China)

机构地区：[1]首都师范大学物理系,北京100048 [2]清华大学精密仪器系,北京100084

出　　处：《科技导报》2025年第5期107-116,共10页Science & Technology Review

基　　金：国家自然科学基金青年科学基金项目(62205173);国家自然科学基金专项项目(62441613)。

摘　　要：计算全息技术是动态三维显示的理想解决方案,具有广阔的应用前景。在当前的技术条件下,计算全息技术面临的最大挑战是全息算法难以同时兼顾计算的精度和速度。为此,提出了卷积误差消除的模型驱动相位型全息图生成网络,实现了高保真相位型全息图的快速生成。首先,研究了角谱法中卷积误差的产生机制,提出了无卷积误差的角谱法,开发了基于无卷积误差角谱法的迭代框架,证实了无卷积误差角谱法对于提升相位型全息图计算精度的有效性;其次,以无卷积误差角谱法作为编码器构建了卷积误差消除的模型驱动相位型全息图生成网络,将相位型全息图的计算时间减小了3个数量级。通过网络生成的相位型全息图,抑制了全息光学重建中的散斑噪声,提高了重建结果的细节质量,平均峰值信噪比高达20.38 dB。伴随着深度梯度显著性和通道效率一致性的继续提升,该网络有望广泛应用在虚拟现实、元宇宙和三维视频通讯等领域。Computer-generated holography is a distinguished and promising choice for dynamic three-dimensional display.Currently,the greatest challenge in computer-generated holography lies in the inability of holographic algorithms to simultaneously achieve both high precision and speed.To address this challenge,a convolution-error-free model-driven neural network is proposed in this work,enabling the fast generation of high-fidelity phase-only computer-generated holograms.Firstly,the mechanism of convolution errors in the angular spectrum method is analyzed.A modified angular spectrum method that eliminates convolution errors is proposed,along with an iterative framework based on this modified method.The effectiveness of the error-free angular spectrum method in enhancing the calculation precision of phase-only holograms is successfully demonstrated.Secondly,a model-driven neural network is developed by incorporating the convolution-error-free angular spectrum method as the decoder.This network achieves a reduction of calculation time for generating phase-only holograms by 3 orders of magnitude.The phase-only holograms generated by this network effectively suppress speckle noise and enhance detail quality in optical reconstructions,achieving an average peak signal-to-noise ratio(PSNR)of 20.38 dB.With ongoing advancements in depth gradient saliency and channel efficiency consistency,the proposed network holds significant potential for widespread application in areas including virtual reality,metaverse,and three-dimensional video communication.

关 键 词：计算全息 三维显示 深度学习 神经网络 卷积误差 

分 类 号：O438.1[机械工程—光学工程]