频谱卷积神经网络实现全息图散斑降噪  

Speckle Noise Reduction of Holograms Based on Spectral Convolutional Neural Network

作　　者：周文静[1] 邹帅 何登科 Hu Jinglu 于瀛洁[1] Zhou Wenjing;Zou Shuai;He Dengke;Hu Jinglu;Yu Yingjie(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China;Graduate School of Information,Product and Systems,Waseda University,Kitakyushu,Fukuoka 8080135,Japa)

机构地区：[1]上海大学机电工程与自动化学院,上海200444 [2]早稻田大学情报生产系统学院,福冈北九州8080135

出　　处：《光学学报》2020年第5期61-68,共8页Acta Optica Sinica

基　　金：国家自然科学基金(61975112,51775326);上海市自然科学基金(18ZR1413700)。

摘　　要：数字全息系统是一种非常先进的成像系统,但相干光源数字全息系统中散斑噪声会对全息图的质量产生不利影响,常规实验降噪或基于传统神经网络算法降噪方法均存在不足。为实现全息图中的散斑降噪以及权衡降噪效率问题,提出一种基于卷积神经网络的单幅全息图快速降噪算法,使用散斑噪声数据集对多等级神经网络进行训练。理论分析及实验结果表明卷积神经网络应用于数字全息图的频谱域去噪能有效提高全息图的质量,且仅使用一幅全息图就可以有效地处理不同等级散斑噪声,在保持去噪性能的前提下,能最大限度保存全息图有效干涉条纹。Digital holographic system is a promising image-forming system, but speckle noise in the coherent light source of digital holographic system adversely affects the quality of holograms. There are some disadvantages in conventional experimental noise reduction or traditional neural network-based noise reduction methods. In order to realize speckle noise reduction in holograms and balance the efficiency of noise reduction, a fast noise reduction algorithm based on convolutional neural network for single hologram is proposed, and the speckle noise dataset is used to train multilevel neural networks. Theoretical analysis and experimental results show that the convolution neural network applied in digital hologram spectrum domain denoising can effectively improve the quality of the hologram, and multilevel speckle noise can be effectively processed by only one hologram. which can save the effective interference fringes of holograms to the maximum extent while maintaining the denoising performance.

关 键 词：数字全息 散斑噪声 频谱降噪 神经网络 

分 类 号：O436[机械工程—光学工程]