物理驱动的中波红外光谱压缩编码与重建

王陆洋; 梁静秋; 赵百轩; 聂海涛; 陈宇鹏; 赵莹泽; 郑凯丰; 秦余欣; 王维彪; 刘钰; 李资政; 吕金光

doi:10.37188/CO.2026-0015

物理驱动的中波红外光谱压缩编码与重建

doi: 10.37188/CO.2026-0015

cstr: 32171.14.CO.2026-0015

1.
中国科学院长春光学精密机械与物理研究所光学系统先进制造全国重点实验室, 吉林长春 130033
2.
中国科学院大学, 北京 100049
3.
中山大学天琴中心 & 物理与天文学院, 国家航天局引力波研究中心, 天琴前沿科学中心, 天琴计划教育部重点实验室, 广东珠海 519082

基金项目: 国家自然科学基金（No. 62575282，No. 62405317，No. 62305339，No. 61805239）；吉林省科技发展计划项目（No. 20260102305JC）；中国科学院青年创新促进会基金（No. 2018254）；吉林省与中国科学院科技合作高技术产业化专项资金项目（No. 2024SYHZ0049）

详细信息

作者简介:
王陆洋（2000—），男，山东东营人，硕士研究生，2023年于哈尔滨工业大学获得学士学位，主要从事中波红外计算光谱成像与压缩感知光谱重建方面的研究。E-mail：wangluyang23@mails.ucas.ac.cn

李资政（1987—），男，吉林长春人，博士，副教授，硕士生导师，2011年于中国科学技术大学获得学士学位，2016年于中国科学院长春光学精密机械与物理研究所获得博士学位，主要从事光学薄膜与纳米结构薄膜材料等方面的研究。E-mail：lizizh@mail.sysu.edu.cn

吕金光（1984—），男，吉林蛟河人，博士，研究员，博士生导师，2008年于吉林大学获得学士学位，2013年于中国科学院长春光学精密机械与物理研究所获得博士学位，主要从事傅里叶变换光谱成像及计算光谱成像等方面的研究。E-mail：jinguanglv@163.com

中图分类号: TP394.1;TH691.9
计量
- 文章访问数: 11
- HTML全文浏览量: 5
- PDF下载量: 0
- 被引次数: 0
出版历程
- 网络出版日期: 2026-06-03

Physics-driven mid-wave infrared spectral compressed encoding and reconstruction

1.
State Key Laboratory of Advanced Manufacturing for Optical Systems, Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, China
2.
University of Chinese Academy of Sciences, Beijing 100049, China
3.
MOE Key Laboratory of TianQin Mission, TianQin Research Center for Gravitational Physics & School of Physics and Astronomy, Frontiers Science Center for TianQin, Gravitational Wave Research Center of CNSA, Sun Yat-sen University (Zhuhai Campus), Zhuhai 519082, China)

Funds: Supported by National Natural Science Foundation of China (No. 62575282, No. 62405317, No. 62305339, No. 61805239); Science and Technology Development Plan Project of Jilin Province (No. 20260102305JC); Youth Innovation Promotion Association of the Chinese Academy of Sciences (No. 2018254); Jilin Province and Chinese Academy of Sciences Science and Technology Cooperation High Tech Special Fund Project (No. 2024SYHZ0049).

More Information

Corresponding author: lizizh@mail.sysu.edu.cn; jinguanglv@163.com

摘要

摘要:
针对现有适配可见光波段的光谱压缩感知重建算法难以实现中波红外尖锐气体吸收光谱高精度重建的问题，本文提出了一种物理驱动的中波红外光谱压缩编码与重建网络架构，旨在实现中波红外尖锐气体吸收光谱的高精度重建。该网络以双分支中波红外光谱重建网络为核心模块，能够分别通过平滑背景重建分支和特征吸收重建分支分别实现平滑背景对数光谱与尖锐气体特征吸光度的精准重建。通过信息融合、物理量转换与全连接层后处理实现中波红外气体吸收光谱的高准确度重建。在对3.7~4.8 μm波段45通道的实际场景气体吸收光谱进行重建的实验中，本文提出的方法达到了峰值信噪比大于28.159 dB、光谱角映射优于0.053 rad的性能指标，对于图像分辨率为320×256的数据立方体重建时间约为0.65 s。该方法有效突破了中波红外光谱高精度重建的技术瓶颈，兼具物理驱动的可解释性与数据驱动的泛化能力，为中波红外压缩感知光谱重建提供了可行技术路径，具有显著的实际应用潜力。
- 中波红外 /
- 光谱成像 /
- 压缩感知 /
- 深度学习
Abstract:
Aiming at the problem that existing spectral compressed sensing algorithms adapted to the visible band are difficult to achieve high-precision reconstruction for sharp gas absorption features in the mid-wave infrared (MWIR) spectra, this paper proposes a physics-driven MWIR spectral compressed encoding and reconstruction network to realize high-precision reconstruction of MWIR spectra with sharp gas absorption features. The dual-branch MWIR spectral reconstruction network serves as the core module of the proposed framework. Specifically, the network consists of two parallel branches, namely the smooth background reconstruction branch and the characteristic absorption reconstruction branch, which respectively realize the accurate reconstruction of smooth background logarithmic spectrum and sharp gas characteristic absorbance. Subsequently, high-accuracy reconstruction of MWIR gas absorption spectra is achieved through information fusion, physical quantity conversion, and post-processing with fully connected layers. Experimental results on the reconstruction of gas absorption spectra within the 3.7−4.8 μm band with 45 channels in real-world scenarios demonstrate that the proposed method achieves a peak signal-to-noise ratio (PSNR) of more than 28.159 dB and a spectral angle mapper (SAM) value of better than 0.053 rad. For a data cube with an image resolution of 320×256, the reconstruction time is approximately 0.65 seconds. This method effectively breaks through the technical bottleneck of high-precision MWIR spectral reconstruction, and it features both the interpretability of physics-driven models and the generalization capability of data-driven models. It provides a feasible technical path for MWIR spectral compressed sensing and exhibits significant potential for practical applications.
- Mid-wave infrared /
- Spectral imaging /
- Compressed sensing /
- Deep learning

HTML全文

图 1 基于多光谱滤光阵列的压缩感知光谱成像原理.(a)光谱成像过程示意图.(b)光谱信号处理流程.

Figure 1. Schematic of the mid-wave infrared spectral imaging using the multispectral filter array. (a) The spectral imaging process. (b) The spectral signal processing workflow.

下载: 全尺寸图片幻灯片

图 2 物理驱动的中波红外光谱压缩编码与重建网络原理

Figure 2. Schematic of the physics-driven mid-wave infrared spectral compressed encoding and reconstruction network

下载: 全尺寸图片幻灯片

图 3 注意力增强的一维残差浓度程长积预测网络结构。(a)网络的总体结构图；(b)多头自注意力模块结构图

Figure 3. Schematic of the attention enhanced 1D residual network for concentration-pathlength product prediction. (a) Schematic of the overall network. (b) Schematic of the multi-head self-attention module.

下载: 全尺寸图片幻灯片

图 4 正则化函数对比图。(a)软阈值正则化函数；(b)硬阈值正则化函数.

Figure 4. Comparison of regularization functions. (a) Regularization function with a soft threshold. (b) Regularization function with a hard threshold.

下载: 全尺寸图片幻灯片

图 5 性能最优滤光片组的光谱透过率

Figure 5. Spectral transmittance of the optimal-performance filters

下载: 全尺寸图片幻灯片

图 6 对503 K背景温度下典型气体吸收光谱的仿真重建结果。(a)二氧化碳(100%)；(b)一氧化二氮(1%)；(c)一氧化碳(10%)；(d)二氧化碳(0.03%)；(e)二氧化硫(10%)；(f)一氧化二氮(0.1%)

Figure 6. Simulated reconstruction results of typical gas absorption spectra at the background temperature of 503 K. (a) Carbon dioxide (100%). (b) Nitrous oxide (1%). (c) Carbon monoxide (10%). (d) Carbon dioxide (0.03%). (e) Sulfur dioxide (10%). (f) Nitrous oxide (0.1%).

下载: 全尺寸图片幻灯片

图 7 对于不同背景温度下目标气体-空气混合体系吸收光谱的仿真重建结果

Figure 7. Simulated reconstruction results of absorption spectra of gas-air mixture systems under different temperature backgrounds

下载: 全尺寸图片幻灯片

图 8 对不同背景温度下纯二氧化碳吸收光谱仿真重建结果。(a) 403 K；(b) 453 K；(c)503 K

Figure 8. Simulated reconstruction results of pure carbon dioxide absorption spectra under different background temperatures. (a) 403 K. (b) 453 K. (c) 503 K.

下载: 全尺寸图片幻灯片

图 9 实验装置图

Figure 9. The image of the experimental setup

下载: 全尺寸图片幻灯片

图 10 不同背景温度下纯二氧化碳吸收光谱的重建结果。(a) 403 K；(b) 453 K；(a) 503 K

Figure 10. Reconstruction results of pure carbon dioxide absorption spectra at different background temperatures. (a) 403 K. (b) 453 K. (a) 503 K.

下载: 全尺寸图片幻灯片

图 11 503 K背景温度下目标气体-空气混合体系吸收光谱的重建结果。(a)二氧化硫(10%)；(b)一氧化二氮(0.1%)；(c) 一氧化二氮(1%)；(d)一氧化碳(10%)

Figure 11. Reconstruction results of absorption spectra of gas-air mixture systems at 503 K background temperature. (a) Sulfur dioxide (10%). (b) Nitrous oxide (0.1%). (c) Nitrous oxide (1%). (d) Carbon monoxide (10%).

下载: 全尺寸图片幻灯片

表 1 前向建模网络的测试结果

Table 1. Testing result of the forward modeling network

指标	平均PSNR (dB)	平均RMSE	平均SAM (rad)
性能	52.713	0.002	0.002

下载: 导出CSV

表 2 构建数据集所用两类吸收体系的气体参数

Table 2. Gas parameters of two types of absorption systems for constructing the dataset

吸收体系类型	目标气体	体积分数(%)	稀释气	目标气体光程(m)	空气柱光程(m)	气压 (atm)
纯目标气体吸收	一氧化二氮	0.1	氮气	0.3	0	1
	一氧化二氮	1
	一氧化碳	10
	二氧化硫	10
	二氧化碳	0.03
	二氧化碳	100	无
目标气体- 空气混合吸收	一氧化二氮	0.1	氮气		0.6
	一氧化二氮	1
	一氧化碳	10
	二氧化硫	10

下载: 导出CSV

表 3 气体吸光度矩阵中含有的气体参数

Table 3. Parameters of the gases in the gas absorbance matrix

气体类型	体积分数(%)	光程 (m)	气压(atm)
一氧化二氮	1	0.3	1
一氧化碳	10
二氧化硫	10
二氧化碳	100
二氧化碳	0.03	0.6

下载: 导出CSV

表 4 本文模型与基础模型的对比

Table 4. Comparison between the model in this article and the basic model

模型	平均PSNR (dB)	平均SAM (rad)	每光谱浮点运算次数(MFLOPs)	参数量
本文模型	46.988	0.010	43.928	16.457 M
PCSED^[10]	33.124	0.049	30.617	4.165 M

下载: 导出CSV

表 5 消融实验结果

Table 5. Results of the ablation experiments

网络处理	平均PSNR (dB)	平均RMSE	平均SAM (rad)
原始模型	46.988	0.004	0.010
无多头注意力	42.865	0.007	0.007
无ECA	43.833	0.006	0.008
无特征吸收重建分支	38.194	0.014	0.097

下载: 导出CSV

表 6 不同背景温度下纯二氧化碳吸收光谱重建性能

Table 6. Reconstruction performance of pure carbon dioxide absorption spectra at different background temperatures

温度	气体类型与浓度	PSNR (dB)	RMSE	SAM (rad)	重建时间(s)
403 K	二氧化碳100%	34.861	0.004	0.029	0.640
453 K		33.528	0.011	0.031	0.643
503 K		30.061	0.031	0.044	0.639

下载: 导出CSV

表 7 503 K背景温度下目标气体-空气混合体系吸收光谱重建性能

Table 7. Reconstruction performance of absorption spectra of gas-air mixture systems at 503 K background temperature

气体类型	气体浓度	PSNR (dB)	RMSE	SAM (rad)	重建时间 (s)
二氧化硫	10%	29.749	0.033	0.040	0.658
一氧化二氮	1%	28.159	0.039	0.053	0.643
一氧化二氮	0.1%	28.194	0.039	0.051	0.664
一氧化碳	10%	28.745	0.033	0.048	0.661

下载: 导出CSV

参考文献(30)

[1]	XING CH ZH, WEI SH C, LI Y K, et al. Fast-hyperspectral imaging remote sensing: emission quantification of NO₂ and SO₂ from marine vessels[J]. Light: Science & Applications, 2025, 14(1): 308.
[2]	KASTEK M, SOSNOWSKI T, ORŻANOWSKI T, et al. Multispectral gas detection method[C]. WIT Transactions on Ecology and the Environment: Vol. 123, WIT Press, 2009: 227-236.
[3]	MERIBOUT M. Gas leak-detection and measurement systems: prospects and future trends[J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 4505813.
[4]	CHEN Y P, LIANG J Q, ZHAO B X, et al. Snapshot infrared Fourier transform imaging spectrometer for transient dynamic sensing[J]. Optics Express, 2025, 33(3): 4024-4043. doi: 10.1364/OE.550328
[5]	吕金光, 梁静秋, 赵百轩, 等. 全景双谱段红外成像干涉光谱测量反演仪器[J]. 中国光学(中英文), 2022, 15(5): 1092-1104. LV J G, LIANG J Q, ZHAO B X, et al. Panoramic bispectral infrared imaging interference spectrum measurement inversion instrument[J]. Chinese Optics, 2022, 15(5): 1092-1104. (in Chinese).
[6]	YANG SH W, QIN H L, YAN X, et al. Mid-wave infrared snapshot compressive spectral imager with deep infrared denoising prior[J]. Remote Sensing, 2023, 15(1): 280. doi: 10.3390/rs15010280
[7]	YUAN L Y, WEN M X, WANG Y M, et al. Development and evaluation of MWIR imaging spectrometer for multi-dimensional detection[J]. Infrared Physics & Technology, 2024, 137: 105148. doi: 10.1016/j.infrared.2024.105148
[8]	孔相金, 李博, 李寒霜, 等. 痕量气体掩星探测高光谱成像光谱仪光学系统设计[J]. 中国光学(中英文), 2024, 17(3): 661-673. KONG X J, LI B, LI H SH, et al. Optical system design of hyperspectral imaging spectrometer for trace gas occultation detection[J]. Chinese Optics, 2024, 17(3): 661-673. (in Chinese).
[9]	ZHANG W Y, SONG H Y, HE X, et al. Deeply learned broadband encoding stochastic hyperspectral imaging[J]. Light: Science & Applications, 2021, 10(1): 108.
[10]	SONG H Y, MA Y G, HAN Y B, et al. Deep-learned broadband encoding stochastic filters for computational spectroscopic instruments[J]. Advanced Theory and Simulations, 2021, 4(3): 2000299. doi: 10.1002/adts.202000299
[11]	WU N F, SUN Y X, HU J T, et al. Intelligent nanophotonics: when machine learning sheds light[J]. eLight, 2025, 5(1): 5. doi: 10.1186/s43593-025-00085-x
[12]	DREMEL J, SCHARF E, RICHTER S, et al. Lensless single-shot multicore fiber endomicroscopy using a single multispectral hologram[J]. Light: Advanced Manufacturing, 2025, 6(4): 27. doi: 10.37188/lam.2025.027
[13]	杨文庆, 霍茨, 孟贺岩, 等. 宽带光谱结构编码计算光谱成像研究进展[J]. 红外与激光工程, 2025, 54(7): 202501. doi: 10.3788/IRLA20250130 YANG W Q, HUO C, MENG H Y, et al. Recent advances in computational spectral imaging based on broadband spectral structure coding[J]. Infrared and Laser Engineering, 2025, 54(7): 202501. (in Chinese). doi: 10.3788/IRLA20250130
[14]	肖树林, 胡长虹, 高路尧, 等. 像元映射变分辨率光谱成像重构[J]. 中国光学(中英文), 2022, 15(5): 1045-1054. doi: 10.37188/CO.2022-0108 XIAO SH L, HU CH H, GAO L Y, et al. Pixel mapping variable-resolution spectral imaging reconstruction[J]. Chinese Optics, 2022, 15(5): 1045-1054. (in Chinese). doi: 10.37188/CO.2022-0108
[15]	YAKO M, YAMAOKA Y, KIYOHARA T, et al. Video-rate hyperspectral camera based on a CMOS-compatible random array of fabry−pérot filters[J]. Nature Photonics, 2023, 17(3): 218-223. doi: 10.1038/s41566-022-01141-5
[16]	XIONG J, CAI X SH, CUI K Y, et al. Dynamic brain spectrum acquired by a real-time ultraspectral imaging chip with reconfigurable metasurfaces[J]. Optica, 2022, 9(5): 461-468. doi: 10.1364/OPTICA.440013
[17]	ARAD B, BEN-SHAHAR O. Sparse recovery of hyperspectral signal from natural RGB images[C]. Proceedings of the 14th European Conference on Computer Vision, Springer, 2016: 19-34.
[18]	徐平, 徐吉, 冯宇超, 等. 面向CASSI的高光谱图像重建算法研究综述(特邀)[J]. 光子学报, 2025, 54(9): 0954206. doi: 10.3788/gzxb20255409.0954206 XU P, XU J, FENG Y CH, et al. A review of hyperspectral image reconstruction algorithms for coded aperture snapshot spectral imaging (invited)[J]. Acta Photonica Sinica, 2025, 54(9): 0954206. (in Chinese). doi: 10.3788/gzxb20255409.0954206
[19]	宋思远, 高博, 王帅, 等. 基于光谱重构的中波红外快照型光谱成像技术研究[J]. 光子学报, 2025, 54(8): 0811001. doi: 10.3788/gzxb20255408.0811001 SONG S Y, GAO B, WANG SH, et al. Mid-wave infrared snapshot spectral imaging via spectral reconstruction[J]. Acta Photonica Sinica, 2025, 54(8): 0811001. (in Chinese). doi: 10.3788/gzxb20255408.0811001
[20]	HUANG L Q, LUO R CH, LIU X, et al. Spectral imaging with deep learning[J]. Light: Science & Applications, 2022, 11(1): 61.
[21]	FENG SH Q, WANG ZH SH, CHENG X B, et al. Superposition Fabry−Perot filter array for a computational hyperspectral camera[J]. Optics Letters, 2023, 48(5): 1156-1159. doi: 10.1364/OL.479622
[22]	YANG J W, CUI K Y, CAI X SH, et al. Ultraspectral imaging based on metasurfaces with freeform shaped meta-atoms[J]. Laser & Photonics Reviews, 2022, 16(7): 2100663. doi: 10.1002/lpor.202100663
[23]	YANG J W, CUI K Y, HUANG Y D, et al. Deep-learning based on-chip rapid spectral imaging with high spatial resolution[J]. Chip, 2023, 2(2): 100045. doi: 10.1016/j.chip.2023.100045
[24]	BIAN L H, WANG ZH, ZHANG Y ZH, et al. A broadband hyperspectral image sensor with high spatio-temporal resolution[J]. Nature, 2024, 635(8037): 73-81. doi: 10.1038/s41586-024-08109-1
[25]	ZHU J P, NIE H T, WANG L Y, et al. Snapshot hyperspectral imaging with high spatial resolution based on transformers[J]. Optics Express, 2025, 33(15): 31731-31755. doi: 10.1364/OE.563614
[26]	张颖, 连彦喆, 张晞, 等. 基于DD-CASSI退化成像模型的高光谱图像重建算法(内封面文章·特邀)[J]. 红外与激光工程, 2026, 55(2): 20250611. doi: 10.3788/IRLA20250611 ZHANG Y, LIAN Y ZH, ZHANG X, et al. Hyperspectral image reconstruction algorithm based on the DD-CASSI degradation imaging model (inner cover paper·invited)[J]. Infrared and Laser Engineering, 2026, 55(2): 20250611. (in Chinese). doi: 10.3788/IRLA20250611
[27]	RÜDINGER A, ATMACA Ö, BAHLINGER V, et al. Bimodal tissue differentiation using hyperspectral imaging and elastographic Fourier transform profilometry[J]. Light: Advanced Manufacturing, 2025, 6(4): 73. doi: 10.37188/lam.2025.073
[28]	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]. Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2018: 7132-7141.
[29]	WANG Q L, WU B G, ZHU P F, et al. ECA-net: efficient channel attention for deep convolutional neural networks[C]. Proceedings of 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2020: 11531-11539.
[30]	ROTHMAN L, TASHKUN S, MIKHAILENKO S, et al. HITRAN on the web[EB/OL]. [2026-01-26]. https://hitran.iao.ru.