pith. sign in

arxiv: 1907.00209 · v1 · pith:DLDFA5KZnew · submitted 2019-06-29 · 📡 eess.IV · cs.CV

High Sensitivity Snapshot Spectrometer Based on Deep Network Unmixing

Xiaoyu Chen , Xu Wang , Lianfa Bai , Jing Han , Zhuang Zhao This is my paper

Pith reviewed 2026-05-25 12:49 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords snapshot spectrometerdeep network unmixingsingle-path designsub-Hadamard codingspectral reconstructionhigh sensitivitylight throughput
0
0 comments X

The pith

A convolutional neural network recovers light intensity from overlapped spectra to enable a compact single-path snapshot spectrometer with higher throughput and SNR.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how a neural network can extract the original light intensity distribution from overlapped dispersive spectra captured in a single optical path. This removes the need for a separate reference path that previous dual-path designs required. The resulting single-path sub-Hadamard spectrometer stays snapshot-capable and high-sensitivity while gaining light throughput. Experiments and simulations indicate the higher throughput produces cleaner reconstructed spectra than the dual-path version. The work therefore trades an added physical path for a learned unmixing step.

Core claim

The authors replace the extra light path of their prior dual-path sub-Hadamard snapshot spectrometer with a convolutional neural network that reconstructs the incident light intensity from the overlapped dispersive spectra. The single-path instrument built on this reconstruction is more compact yet preserves snapshot operation and high sensitivity, and it delivers higher signal-to-noise ratio spectra because all available light reaches the detector without division between paths.

What carries the argument

Convolutional neural network that unmixes overlapped dispersive spectra to recover the original light intensity distribution for subsequent spectrum reconstruction.

If this is right

  • The instrument becomes physically smaller while retaining snapshot and high-sensitivity operation.
  • All incident light contributes to the measurement instead of being split, raising throughput.
  • Reconstructed spectra exhibit higher signal-to-noise ratio than those from the dual-path predecessor.
  • Simulated and experimental comparisons confirm the SNR advantage holds across multiple test cases.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same unmixing step could substitute for auxiliary reference paths in other coded-aperture or dispersive instruments.
  • Portability improves because the optical train is shortened, which may suit field-deployed spectral sensing.
  • Performance under low-light or rapidly changing scenes would test whether the network remains accurate outside the training distribution.

Load-bearing premise

The neural network reconstructs the true light intensity distribution from the overlapped spectra without adding systematic errors that would offset the throughput gain.

What would settle it

An experiment in which the network-recovered spectrum shows larger deviation from a known ground-truth source than the dual-path measurement, after accounting for the measured throughput difference.

Figures

Figures reproduced from arXiv: 1907.00209 by Jing Han, Lianfa Bai, Xiaoyu Chen, Xu Wang, Zhuang Zhao.

Figure 1
Figure 1. Figure 1: Principle of Hadamard transform spectrum measurement [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Principle difference between traditional HTS and sub-Hadamard snapshot spectrometer In the traditional method, if we want to obtain the light intensity, we need an extra imaging path to capture the light intensity. However, as mentioned above, there are many problems in the dual-path scheme. Therefore, an appropriate method is needed to obtain the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Structural sketch of network model The unmixing module simulates the inverse process of superimposed and extracts the primary features through spectral direction convolution and deconvolution. In the enhance module, the output feature map is further enhanced to supplement more details such as image contrast, brightness, texture features and so on. The mean square error function is used as the [PITH_FULL_I… view at source ↗
Figure 4
Figure 4. Figure 4: non 4. SNR ana In order to an the traditiona sub-Hadamard between the reconstructed [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hyperspectral data from lesun[25]. (a)The imag [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Reconstructed results. From left to right, they ar [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Reco dispersive imag (400-500nm) to As show the original i becomes more In order reconstructed light intensity experiment, w ment (TI). The ating uses THO ze is 13.68um onstructed results o ge, and the reconst truncate the spectr n in the Figure image well. H e serious, and t to quantitativ spectra, we c y, the sub-Hada we ignore the li e camera used ORLABS GT2 × 13.68um, an Figur of actual experime tructe… view at source ↗
Figure 9
Figure 9. Figure 9: Spectral reconstruction simulation results. SNRN [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of the two-path scheme with the proposed method in the same case. SNRNet=15.1712dB, SNRsub-S=12.4685dB At the same time, as shown in [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: The results of actual experiments, (a) Spectral images. (b) Reconstruction results [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: 30dB gain experimental results. SNRNet=14.0912dB, SNRslit=6.2915dB, SNRsub-S=13.052dB It can be seen that in the actual test, our scheme has a better SNR than the dual-path scheme because of the advantage of light throughput. 6. Conclusion On the basis of sub-Hadamard snapshot spectral detection, this paper proposes an innovative method of spectral image unmixing using a convolutional neural network to ob… view at source ↗
read the original abstract

In this paper, we present a convolution neural network based method to recover the light intensity distribution from the overlapped dispersive spectra instead of adding an extra light path to capture it directly for the first time. Then, we construct a single-path sub-Hadamard snapshot spectrometer based on our previous dual-path snapshot spectrometer. In the proposed single-path spectrometer, we use the reconstructed light intensity as the original light intensity and recover high signal-to-noise ratio spectra successfully. Compared with dual-path snapshot spectrometer, the network based single-path spectrometer has a more compact structure and maintains snapshot and high sensitivity. Abundant simulated and experimental results have demonstrated that the proposed method can obtain a better reconstructed signal-to-noise ratio spectrum than the dual-path sub-Hadamard spectrometer because of its higher light throughput.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a convolutional neural network (CNN) to recover the original light intensity distribution from overlapped dispersive spectra in a single-path sub-Hadamard snapshot spectrometer. This replaces the second optical path used in the authors' prior dual-path design, yielding a more compact instrument that retains snapshot capability and high sensitivity. The central claim is that the single-path version achieves higher reconstructed SNR than the dual-path version because of greater light throughput, with the CNN unmixing step presented as a direct substitute for direct measurement; this is supported by simulated and experimental results.

Significance. If the CNN recovery step can be shown to introduce negligible systematic bias relative to the photon-noise floor, the work would demonstrate a practical route to higher-throughput snapshot spectrometers without added hardware paths. The provision of both simulated and experimental demonstrations is a positive feature, as is the explicit comparison to the authors' own prior dual-path instrument.

major comments (2)
  1. [Results section (experimental SNR comparison)] Results section (experimental SNR comparison): the reported improvement in reconstructed spectrum SNR is attributed to higher light throughput enabled by the single-path design, yet no isolated metric (e.g., pixel-wise MSE, correlation, or residual map) is provided that compares the CNN-recovered intensity distribution against a direct measurement of the same distribution; without this separation, the SNR gain cannot be unambiguously assigned to throughput rather than to favorable simulation conditions or post-processing.
  2. [Methods section (network training and validation)] Methods section (network training and validation): the claim that the CNN accurately reconstructs the true intensity distribution 'instead of adding an extra light path' requires a held-out test set that quantifies reconstruction fidelity on intensity maps independent of the final spectrum recovery; the absence of such a metric leaves the central substitution argument untested at the load-bearing step.
minor comments (2)
  1. [Abstract and Introduction] The abstract and introduction use the phrase 'for the first time' without a supporting literature comparison that distinguishes the present unmixing task from prior CNN-based spectral unmixing work.
  2. [Figure captions] Figure captions for the experimental results should explicitly state the number of independent trials and the precise definition of 'reconstructed SNR' (e.g., whether it is per-wavelength or integrated).

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive comments. Below we respond point-by-point to the two major comments, indicating where revisions will be made and where experimental constraints limit what can be provided.

read point-by-point responses

  1. Referee: Results section (experimental SNR comparison): the reported improvement in reconstructed spectrum SNR is attributed to higher light throughput enabled by the single-path design, yet no isolated metric (e.g., pixel-wise MSE, correlation, or residual map) is provided that compares the CNN-recovered intensity distribution against a direct measurement of the same distribution; without this separation, the SNR gain cannot be unambiguously assigned to throughput rather than to favorable simulation conditions or post-processing.

    Authors: We agree that an isolated metric isolating CNN reconstruction error would strengthen attribution of the SNR gain. In simulations, ground-truth intensity maps are known; the CNN achieves low pixel-wise MSE and high correlation with these maps, with residuals well below the photon-noise floor. We will add residual maps and quantitative intensity-error metrics to the revised Results section. In the experimental single-path configuration, however, no direct measurement of the intensity distribution exists, because that measurement would require the second optical path the design eliminates. The reported experimental SNR improvement is therefore measured on the final spectra and is consistent with the measured doubling of collected light relative to the dual-path instrument. revision: partial

  2. Referee: Methods section (network training and validation): the claim that the CNN accurately reconstructs the true intensity distribution 'instead of adding an extra light path' requires a held-out test set that quantifies reconstruction fidelity on intensity maps independent of the final spectrum recovery; the absence of such a metric leaves the central substitution argument untested at the load-bearing step.

    Authors: The network was trained on simulated intensity-to-spectrum pairs, with a held-out validation subset used during training to monitor convergence. Separate quantitative fidelity metrics on intensity maps (MSE, correlation) from an independent test set were not reported. We accept that an explicit demonstration of intensity-map fidelity would better support the substitution claim and will add a dedicated evaluation subsection and table in the Methods section of the revision. revision: yes

standing simulated objections not resolved
  • Direct experimental comparison of CNN-recovered intensity maps against a measured ground-truth intensity distribution cannot be performed, because obtaining that ground truth requires the dual-path hardware the single-path design removes.

Circularity Check

0 steps flagged

Minor self-citation to prior dual-path design; CNN recovery and SNR claims rest on simulation/experiment rather than definitional reduction

full rationale

The paper references its own prior dual-path spectrometer only to motivate the single-path construction and then introduces a new CNN unmixing step whose output is validated by simulated and experimental SNR results. No equations, fitted parameters, or self-citation chains are shown that make the recovered intensity or final spectrum equivalent to the input by construction. The throughput advantage is presented as an empirical outcome of the optical change plus network recovery, not a tautology. This is the normal non-circular case for a methods paper that reports external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no free parameters, axioms, or invented entities are stated or derivable from the provided text.

pith-pipeline@v0.9.0 · 5661 in / 1072 out tokens · 29128 ms · 2026-05-25T12:49:41.545035+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    Z. Zhao, L. Bai, J. Han and J. Yue, High-SNR snapshot multiplex spectrometer with sub-Hadamard-S matrix coding. arXiv preprint arXiv:1905.07543, 2019

    work page arXiv 1905

  2. [2]

    Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,

    E. J. Candes, J. Romberg, T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information,” IEEE Trans. on Information Theory 52, 489-509(2006)

    work page 2006

  3. [3]

    Compressed sensing,

    D. L. Donoho, “Compressed sensing,” IEEE Trans. on Information Theory 52, 1289-1306(2006)

    work page 2006

  4. [4]

    Compressive imaging spectrometers using coded apertures,

    D. J. Brady, M. E. Gehm, “Compressive imaging spectrometers using coded apertures,” Visual Information Processing 6246, 62460A-1-62460A-9(2006)

    work page 2006

  5. [5]

    S i n g l e - s h o t compressive spectral imaging with a dual-disperser architecture,

    M . E . G e h m , R . J o h n , D . J . B r a d y , R . M . W i l l e t t a n d T . J . S c h u l z , “ S i n g l e - s h o t compressive spectral imaging with a dual-disperser architecture,” Opt. Express 15, 14013-14027(2007)

    work page 2007

  6. [6]

    Single disperser design for coded aperture snapshot spectral imaging,

    A. Wagadarikar, R. John, R. Willett and D. J. Brady, “Single disperser design for coded aperture snapshot spectral imaging,” Appl. Opt 47, B44-B51(2008)

    work page 2008

  7. [7]

    Backtracking-based matching pursuit method for sparse signal reconstruction,

    H. Huang, A. Makur, “Backtracking-based matching pursuit method for sparse signal reconstruction,” IEEE Signal Processing Letters 18, 391-394(2011)

    work page 2011

  8. [8]

    Sparse solution of underdetermined linear equations by stagewise orthogonal matching pursuit,

    D. Donoho, Y . Tsaig, I. Drori and J. L. Starck, “Sparse solution of underdetermined linear equations by stagewise orthogonal matching pursuit,” Department of Statistics, Stanford University, Technical Report, 2006

    work page 2006

  9. [9]

    Cosamp: Iterat ive signal recovery from incomplete and inaccurate samples,

    D. Needell, J. Tropp, “Cosamp: Iterat ive signal recovery from incomplete and inaccurate samples,” Applied and Computational Harmonic Analysis 26, 301-321(2009)

    work page 2009

  10. [10]

    The in-crowd algorithm for fast basis pursuit denoising,

    P. Gill, A. Wang, A. Molnar, “The in-crowd algorithm for fast basis pursuit denoising,” IEEE Trans. on Signal Processing 59, 4595-4605(2011)

    work page 2011

  11. [11]

    An iterative thresholding algorithm forlinear inverse problems with a sparsity constraint,

    I. Daubechies, M. Defrise, C. De-Mol, “An iterative thresholding algorithm forlinear inverse problems with a sparsity constraint,” Communications on Pure and Applied Mathematics 57, 1413-1457(2004)

    work page 2004

  12. [12]

    A new TwIST: two-step iterative shrinking/thresholding algorithms for image restoration,

    J. M. Bioucas-Dias, M. A. T. Figueiredo, “A new TwIST: two-step iterative shrinking/thresholding algorithms for image restoration,” IEEE Trans. on Image Processing 16, 2992-3004(2007)

    work page 2007

  13. [13]

    Iteratively reweighted algorithms for compressive sensing,

    R. Chartrand, W. Yin, “Iteratively reweighted algorithms for compressive sensing,” Proceedings of IEEE Intern ational Conference on Acous tics, Speech, and Signal Processing, 3869-3872(2008)

    work page 2008

  14. [14]

    Sparse Bayesian learning and the relevance vector machine,

    M. E. Tipping, “Sparse Bayesian learning and the relevance vector machine,” Journal of Machine Learning Research 1, 211-244(2001)

    work page 2001

  15. [15]

    Fast marginal likelihood maximization for sparse Bayesian models,

    M. E. Tipping, A. C. Faul, “Fast marginal likelihood maximization for sparse Bayesian models,” Proceedings of the 9th International Workshop on Artificial Intelligence and Statistics, Key West, FL, 3-6(Jan.2003)

    work page 2003

  16. [16]

    Fast bayesian matching pursuit,

    P. Schniter, L. C. Potter, J. Ziniel, “Fast bayesian matching pursuit,” in Workshop on Information Theory and Applications, La Jolla, CA, Jan.2008

    work page 2008

  17. [17]

    Performance comparison of aperture codes for multimodal, multiplex spectroscopy,

    A. Wagadarikar, M. E. Gehm and D. J. Brady, “Performance comparison of aperture codes for multimodal, multiplex spectroscopy,” Appl. Opt 46, 4932-4942(2007)

    work page 2007

  18. [18]

    Static two-dimensional aperture coding for multimodal, multiplex spectroscopy,

    M . E . G e h m , S . T . M c C a i n , N . P . P i t s i a n i s , D . J . B r a d y , P . P o t u l u r i , a n d M . E . Sullivan, “Static two-dimensional aperture coding for multimodal, multiplex spectroscopy,” Appl. Opt 45, 2965(2006)

    work page 2006

  19. [19]

    Longwave infrared (LWIR) coded aperture dispersive spectrometer,

    C. Fernandez, B. D. Guenther, M. E. Ge hm, D. J. Brady, and M. E. Sullivan, “Longwave infrared (LWIR) coded aperture dispersive spectrometer,” Opt. Express 15, 5742-5753(2007)

    work page 2007

  20. [20]

    Denoising analysis of Hadamard transform spectrometry,

    Y u e . J , H a n . J , Z h a n g . Y a n d B a i . L , “ Denoising analysis of Hadamard transform spectrometry,” Opt. Letters 39, 3744-3747(2014)

    work page 2014

  21. [21]

    Denoising analysis of spatial pixel multiplex coded spectrometer with Hadamard H-matrix,

    Yue. J, Han. J, Li. L and Bai. L, “Denoising analysis of spatial pixel multiplex coded spectrometer with Hadamard H-matrix,” Optics Communications 407, 355-360(2018)

    work page 2018

  22. [22]

    Z. Wu, C. Shen, and A. van den Hengel. High-performance Semantic Segmentation Using Very Deep Fully Convolutional Networks. arXiv e-prints, page arXiv:1604.04339, Apr 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  23. [23]

    ERFNet: Efficient Residual Factorized ConvNet for Real-Time Semantic Se gmentation,

    E. Romera, J. M. Alvarez, L. M. Bergasa and R. Arroyo, “ERFNet: Efficient Residual Factorized ConvNet for Real-Time Semantic Se gmentation,” IEEE Trans. on Intelligent Transportation Systems 19, 263-272(2018)

    work page 2018

  24. [24]

    Multiframe image estimation for coded aperture snapshot spectral imagers,

    D. Kittle, K. Choi, A. Wagadarikar and D. J. Brady, “Multiframe image estimation for coded aperture snapshot spectral imagers,” Appl. Opt 49, 6824-6833(2010)

    work page 2010

  25. [25]

    Indian Pines, http://lesun.weebly.com/hyperspectral-data-set.html

    work page

  26. [26]

    Chakrabarti and T

    A. Chakrabarti and T. Zickler. Statistics of Real-World Hyperspectral Images. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 193– 200(2011)

    work page 2011