高空间分辨率高可见度的太赫兹光谱成像研究  

Terahertz Spectral Imaging With High Spatial Resolution and High Visibility

作　　者：褚致弘 张逸竹 曲秋红 赵晋武 何明霞[1,2] CHU Zhi-hong;ZHANG Yi-zhu;QU Qiu-hong;ZHAO Jin-wu;HE Ming-xia(State Key Laboratory of Precision Measuring Technology and Instruments,Tianjin University,Tianjin 300072,China;School of Precision Instrument and Optoelectronics Engineering,Tianjin University,Tianjin 300072,China;LET Terahertz(Tianjin)Technology Co.,Ltd.,Tianjin 300019,China)

机构地区：[1]天津大学精密测试技术及仪器国家重点实验室,天津300072 [2]天津大学精密仪器与光电子工程学院,天津300072 [3]莱仪特太赫兹(天津)科技有限公司,天津300019

出　　处：《光谱学与光谱分析》2023年第2期356-362,共7页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(61675151,12174284);海南省科技重大专项(ZDKJ2019013)资助。

摘　　要：太赫兹光谱成像,不但包括在二维图像空间的强度信息,同时可以得到太赫兹波段的光谱信息,构成了一个三维的数据矩阵。由于受到太赫兹成像系统内部硬件的限制和影响,太赫兹频域较高频段处信号存在能量弱、信噪比低的特点,导致所成的太赫兹图像普遍存在分辨率低、对比度低等问题。因此,利用三维数据矩阵,应用适合的算法,实现了提高太赫兹光谱成像空间分辨率、边缘细节可见度的目的。搭建了三维可移动式太赫兹时域光谱成像系统,实现了对标准高分辨率板的二维扫描。对该系统所采集到的信号分别进行时域、频域等多种方式成像对比,结合瑞利判据和分辨率标尺对成像系统的空间分辨率、景深进行标定,研究了提高太赫兹光谱成像的空间分辨率算法。然后,针对太赫兹频域高频区域信噪比低、对比度低、噪声原因复杂的特点,结合深度残差学习的图像去噪理论,提出了太赫兹图像深度去噪网络,在训练集中引入成像系统中真实的“太赫兹残差噪声”。最后,利用所训练出的模型对太赫兹频域高频区域图像进行盲去噪,并用重建图像分别与原始成像结果和传统太赫兹去噪算法结果进行比较,分别从主观和客观两个方面评价了不同算法对太赫兹频域高频图像的去噪效果。实验结果表明,通过该算法实现了极限空间分辨率约为157μm,去噪后图像极限空间分辨率处的瑞利判据鞍-峰比约为0.623,图像整体对比度为46.635;空间分辨率相比传统成像方法提高了约一倍,对比度提高约26%。研究结果为高空间分辨率高可见度的太赫兹光谱成像方式提供了一种新的规范,并针对太赫兹频域较高频区域的图像噪声问题提供了一种新的解决方案。Terahertz spectral imaging not only includes the intensity information in two-dimensional image space but can also obtain spectral information in the terahertz band,constituting a three-dimensional data matrix.Due to the limitation and influence of the internal hardware of the Terahertz imaging system,the signals in the higher frequency band of the terahertz frequency domain have weak energy and low signal-to-noise ratio,resulting in the problems of low resolution and low contrast of the terahertz images.Therefore,this paper improves the spatial resolution and edge detail visibility of terahertz spectral imaging by using a three-dimensional data matrix and a suitable algorithm.In this paper,a three-dimensional portable Terahertz time-domain spectroscopy imaging system is built to realize the two-dimensional scanning of standard high-resolution plates.The signals collected by the system were compared in the time domain and frequency domain,respectively.The spatial resolution and depth of field of the imaging system were calibrated by combining the Rayleigh criterion and resolution scale,and the spatial resolution algorithm for improving THZ spectral imaging was studied.Then,aiming at the characteristics of low SNR,low contrast and complex noise causes in the high-frequency region of the Terahertz frequency domain,combined with the image denoising theory of deep residual learning,a terahertz image depth denoising network is proposed,which introduces the real“terahertz residual noise”in the imaging system in the training set.Finally,the reconstructed images are compared with the original images and the traditional terahertz denoising algorithm results.The denoising effects of different algorithms on the high-frequency images in the terahertz frequency domain are evaluated from subjective and objective aspects.Experimental results show that the limit spatial resolution of the proposed algorithm is about 157μm,the saddle-peak ratio of the Rayleigh criterion at the limit spatial resolution of the denoised image is

关 键 词：太赫兹光谱成像 空间分辨率 图像去噪 深度卷积神经网络 

分 类 号：O433[机械工程—光学工程]