摘要
太赫兹光谱成像技术不仅能获取检测对象在三维图像空间中的几何信息,同时也能提取其在太赫兹波段的光谱信息,因此该技术已在众多领域表现出巨大的应用潜力.但由于技术设备结构复杂,成本较高,且太赫兹波长较长导致其成像空间分辨率较低,边缘细节模糊,如何在现有设备基础上提高太赫兹成像分辨率成为目前亟需解决的关键问题.针对该问题,以STC89C51单片机为检测对象,利用太赫兹时域光谱数据,结合深度学习,实现了太赫兹光谱成像的超分辨率重建.试验采用太赫兹时域光谱检测系统对样品进行逐点扫描,指定0.738 THz进行频域成像,通过构建点扩散函数对太赫兹图像去卷积增强,并确定以光束穿透深度z=2 mm的去卷积太赫兹图像作为参考图像.考虑太赫兹图像实际采集过程中可能受到多种复杂噪声的影响,通过双三次插值降采样(BI)、高斯模糊下采样(BD)、双三次插值下采样+高斯白噪声(BN)与高斯模糊下采样+高斯白噪声(DN)4种不同的降采样方式模拟太赫兹降质图像.采用Real-ESRGAN深度学习方法对降质图像进行超分辨率重建,同时与SRResNet、EDSR、SRGAN、ESRGAN方法的重建结果进行对比.采用峰值信噪比、结构相似性以及主观平均得分3种评价指标对重建结果进行评价.评价结果表明:Real-ESRGAN的BI、BD、BN、DN 4种降质图像的超分辨率重建效果的各项指标表现均优于其他算法,实现了对芯片特征信息的增强和图像成像精度的提高,为太赫兹图像超分辨率重建技术提供了一种新的优化思路.
Terahertz spectral imaging technology can obtain the geometric information of the detected object in three-dimensional image space and spectral information in the terahertz band as well,so it has shown huge application potential in many fields.But terahertz spectroscopy equipment has a complex structure and high cost,moreover,due to the long terahertz wavelength,its imaging spatial resolution is low and edge details are blurry,how to improve the resolution of terahertz imaging based on existing equipment has become a key issue that urgently needs to be solved.To address these problems,this paper,taking the STC89C51 microcontroller as the detection object,used terahertz time-domain spectral data and combined with deep learning to achieve super-resolution reconstruction of terahertz spectral imaging.In the experiment a terahertz time domain spectral detection system was used to scan the samples point by point and 0.738 THz was set for frequency domain imaging.The point spread function was established to enhance the terahertz image by deconvolution,and the deconvolution terahertz image with beam penetration depth z=2 mm was determined to be the reference image.Considering that the actual acquisition process of terahertz images can be affected by various complex noises,this paper simulated degraded terahertz images by four different down-sampling methods including bicubic interpolation downsampling(BI),Gaussian fuzzy downsampling(BD),bicubic interpolation downsampling+Gaussian white noise(BN),and Gaussian fuzzy downsampling+Gaussian white noise(DN).Five deep learning methods such as SRResNet,EDSR,SRGAN,ESRGAN,and Real-ESRGAN were used to perform super-resolution reconstruction of degraded images,and compare and analyze the reconstruction results.The reconstruction results were evaluated based on three indicators:peak signal-to-noise ratio,structural similarity,and average subjective score.According to the experiment,the super-resolution reconstruction effect of the degraded images generated by BI,BD,BN and DN outperf
作者
邹佳岐
周胜灵
祝诗平
李博鑫
唐茂杰
张越
刘寅峰
ZOU Jiaqi;ZHOU Shengling;ZHU Shiping;LI Boxin;TANG Maojie;ZHANG Yue;LIU Yinfeng(College of Engineering and Technology,Southwest University,Chongqing 400715,China)
出处
《西南大学学报（自然科学版）》
CAS
CSCD
北大核心
2024年第10期200-211,共12页
Journal of Southwest University(Natural Science Edition)
基金
国家自然科学基金面上项目(62005227,31771670)
重庆市自然科学基金面上项目(cstc2020jcyj-msxmX0300)。
关键词
太赫兹光谱
点扩散函数
图像处理
深度学习
超分辨率
terahertz spectroscopy
point spread function
image processing
deep learning
super-resolution
