太赫兹时域光谱成像增强算法  

Enhancement Algorithm of Terahertz-Time-Domain Spectral Imaging

作　　者：王志勇[1] 赵浩男 陈柏彤 Wang Zhiyong;Zhao Haonan;Chen Baitong(School of Mechanical Engineering,Tianjin University,Tianjin 300350,China)

机构地区：[1]天津大学机械工程学院,天津300350

出　　处：《天津大学学报（自然科学与工程技术版）》2024年第8期779-786,共8页Journal of Tianjin University：Science and Technology

基　　金：国家自然科学基金资助项目(12041201,12072229,12172251).

摘　　要：低分辨率是扫描太赫兹时域光谱(THz-TDS)系统图像的一个重要缺点.本文提出了一种针对太赫兹时域光谱图像的超分辨率增强算法,该算法结合了多层感知机(MLP)和超分辨率卷积神经网络(SRCNN),创建了一个复合网络.本文算法针对太赫兹光谱图像的特点,通过引入新的目标函数,避免了传统机器学习算法需要采集或模拟生成大量训练集的弊端,实现了训练图像即目标图像的单幅光谱图像增强.为了实现这一目标,本文算法的基本原理是,让试件产生一个刚体位移,利用THz-TDS系统采集位移前后两幅三维(1个时间维,2个空间维)光谱图像作为输入数据.机器学习网络包括两部分:首先,利用一个MLP网络实现三维光谱图像到二维光强图像的转化;其次,采用传统针对二维图像的SRCNN网络获取一幅高分辨率图像,对位移前后图像处理后计算得到新的高分辨率图像的位移场,并将位移场方差作为目标函数,再通过机器学习算法,优化网络中的成像参数,实现太赫兹光谱图像的分辨率增强.典型验证性实验最终得到的峰值信噪比为42.65 dB,结构相似度为0.816,均比其他现有方法高,表明本文算法能获得良好的图像增强.Low resolution is a significant disadvantage of images obtained by scanning terahertz-time-domain spectroscopy(THz-TDS)systems.In this study,a super-resolution enhancement algorithm for THz-TDS is proposed.The proposed algorithm is a combination of a multilayer perceptron(MLP)and a super-resolution convolutional neural network(SRCNN),creating a composite network.We introduce an objective function,wherein the algorithm avoids the disadvantages of traditional machine learning algorithm that must collect or simulate several training sets,and realizes single spectral image enhancement of the training image,i.e.,the target image.To achieve this goal,the basic principle of the algorithm is to generate a rigid body displacement of the specimen and use the THz-TDS system to collect two three-dimensional(3D)(one time dimension and two space dimensions)spectral images before and after the displacement and consider them as input data.A machine learning network has two parts.First,an MLP network is used to convert the 3D spectral image to a 2D light-intensity image.Second,the traditional SRCNN network is used to obtain a high-resolution image.The images before and after the displacement are processed,and the displacement field of the new high-resolution image is calculated.The displacement field variance is considered as the objective function.Finally,the image parameters in the network are optimized using a machine learning algorithm to achieve the resolution enhancement of terahertz spectral images.The effectiveness of the proposed algorithm is proved by performing typical validation experiments.The final result is a peak signal-to-noise ratio of 42.65 dB and a struc-tural similarity of 0.816,which is higher than other existing methods,indicating that this study can achieve just the right image enhancement.

关 键 词：太赫兹时域光谱(THz-TDS) 光谱信息 扫描成像 图像增强 

分 类 号：O433[机械工程—光学工程]