基于深层卷积神经网络的连续光空间外差干涉图降噪方法  

Denoising method of continuous light spatial heterodyne interferogram based on deep convolutional neural network

作　　者：罗炜 叶松[1,2] 熊伟 张紫杨[1,2] 王新强 李树 王方原[1,2] LUO Wei;YE Song;XIONG Wei;ZHANG Ziyang;WANG Xinqiang;LI Shu;WANG Fangyuan(School of Optoelectronic Engineering,Guilin University of Electronic Technology,Guilin 541004,China;Guangxi Key Laboratory of Optoelectronic Information Processing,Guilin 541004,China;Anhui Province Key Laboratory of Optical Quantitative Remote Sensing,Hefei Institutes of Physical Science,Chinese Academy of Sciences,Hefei 230031,China;College of Environmental Science and Optoelectronic Technology,University of Science and Technology of China,Hefei 230031,China)

机构地区：[1]桂林电子科技大学光电工程学院,广西桂林541004 [2]广西光电信息处理重点实验室,广西桂林541004 [3]中国科学院合肥物质科学研究院光学定量遥感安徽省重点实验室,安徽合肥230031 [4]中国科学技术大学环境科学与光电技术学院,安徽合肥230031

出　　处：《红外与激光工程》2025年第3期352-365,共14页Infrared and Laser Engineering

基　　金：国家重点研发计划项目(2022YFB3901800,2022YFB3901803);桂林电子科技大学研究生教育创新计划项目(2024YCXB14)。

摘　　要：空间外差光谱仪在探测过程中可能存在着噪声干扰的现象,导致具有连续光谱的目标特征信号被掩盖而无法获取所需信息,因此需要有效方法降低噪声对连续光空间外差干涉图的影响。提出了一种基于深层卷积神经网络的连续光空间外差干涉图降噪方法,利用深层卷积结合残差的方式去除高斯噪声。结果表明该方法可以对连续光空间外差干涉图有效地去噪并复原目标光谱信号。在Sigma=25高斯噪声条件下,深层卷积神经网络的降噪干涉图在峰值信噪比、结构相似性和光谱差值方面分别达到51.74 dB、0.9997和2.20%,优于其他算法最佳值2.83 dB、0.001和0.49%。进一步的研究还表明,深层卷积神经网络的层数对降噪性能有重要影响,为网络模型的优化提供了有益的参考。最后将该方法应用在“高分五号”数据中,网络模型表现出了不错的降噪有效性。该工作在降低噪声对连续光空间外差光谱信息的影响和实现对目标的高精度探测等方面具有研究意义和应用价值。Objective Spatial heterodyne spectroscopy is a novel Fourier transform interferometric spectroscopy technique that achieves high spectral resolution within a specific wavelength range.It has been widely used in fields like atmospheric remote sensing,astronomical observation,and mineral detection.When applied to remote sensing of targets with continuous spectra,the resulting spatial heterodyne interferogram displays only a few interference fringes,yet contains rich spectral information.However,due to interference from complex environments and electronic components,the spatial heterodyne spectrometer may encounter noisy signals during target detection,leading to the destruction of the collected interference fringes.As a result,the spectral features are obscured by various types of noise,making it difficult to obtain valuable research data during the inversion process.As remote sensing research advances,effective methods to reduce the impact of noise on the information contained in continuous light spatial heterodyne interferograms are increasingly needed.Methods Based on the principles of spatial heterodyne spectroscopy,a deep convolutional neural network is constructed using a residual learning approach to predict Gaussian noise,combined with batch normalization to accelerate training and improve network performance(Fig.2).By subtracting the predicted Gaussian noise from the noisy continuous light spatial heterodyne interferogram,the corresponding denoised interferogram is obtained.The effectiveness and superiority of this method are validated through comparisons with other denoising methods using visual effects,PSNR,SSIM,and spectral differences.By comparing the denoising performance of deep convolutional neural networks with different numbers of layers,the study provides effective suggestions for constructing and optimizing the network architecture.Results and Discussions The SHI-DnCNN(Spatial heterodyne interferograms-Denoise convolutional neural networks)was constructed and trained,which is capable of denoisin

关 键 词：空间外差光谱 降噪 神经网络 干涉图 连续光 

分 类 号：O433.4[机械工程—光学工程]