An Efficient and Portable LED Multispectral Imaging System and Its Application to Human Tongue Detection

Round 1

Reviewer 1 Report

While it is probably obvious for those familiar with this area, the cosine similarity used to group pixels could be more clearly presented so those who are not familiar might have a better chance to understand.

Figure 2 shows the spectral response of the LEDs that are used for imaging and this spectral response is used to guide the reconstruction algorithm. It is clear from the spectral response that there is no illumination from the LEDs at 400 nm however, Figure 5 presents reconstructed spectra with spectral information at this wavelength. Please clarify how reflected light at this wavelength is detected when it not present in the initial illumination.

Author Response

We greatly appreciate the reviewer’s comments. We have made revisions accordingly.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

I am not an expert in exactly this field but I read the paper and tried to understand. I guess I can say that the paper is coherent in itself and sounds reasonable. It is, however, hard to understand due to English difficulties. 

Almost every sentence is affected. As an example, in the first sentence of the conclusions it must be (i) "Compared to the" instead of "Comparing to the", (ii) "is cost-effective" instead of "is costeffectiveness", (iii) "easy to implement" instead of "easy to implementation". So, the overoall mistake is that adverbs (how something is done) become nouns.

Hopefully, this hint helps a bit. I recommend first to improve the English language before continuing. All the best.

Author Response

We greatly appreciate the reviewer’s comments. We have made revisions accordingly.

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Main comments:

The authors introduce a novel multispectral system, based on ring LEDs and an RGB camera, thus obtaining 5 optical channels. With the purpose to improve the spectral resolution (from 5 to 31 spectral bands), the authors use a spectral retrieval algorithm (proposed by Perkkinen et. al., 1989),  and give a solution to improve the time-consuming processing, to achieve a 31-bands multispectral cube. The main constraint in this work is, precisely, the use of an ancient algorithm, which has been overpassed by others mentioned by the authors (such as the ones based on machine learning, compressing sampling, and others based on PCA, ICA, or NNMF [1-3]). Such new algorithms provide fast and accurate results to the spectral retrieval procedure. Thus, the problem stated by the authors is questionable. The authors indicate (lines 137-137) that the algorithm proposed by Perkkinen et. al. takes more than 2h to achieve a result. However, after reading the article of Perkkinen, this data is not provided. Due to the problem solved by the authors is not well justified, and then not relevant, we propose to reject its publication.

General comments:

- The algorithm proposed by Perkkinen shows a spectral retrieval based on the orthogonal basis functions calculated from a spectral database of a Munsell color chart. So, the algorithm works based on a-priori spectral information of a target, and to be used on the same targets. Did the authors probe with a spectral database of tongue spectral reflectances? Of course, a proper spectral database provides enough information in order to improve the retrieval process. This way should be explored by the authors.

- lines 90-91: the authors should provide more consistent information to choose the CWL (center wavelength) and spectral shape of the LEDs. For example, are the emitted spectra orthogonal? Which is the effect of choosing other LEDs?
- line 101:  The Eq.1 is confusing, due to the term "I" being indicated such as the response of the 5 sensors (so, the intensity of the 5 gray images). However,  the spectral dimension is L=31 bands. Clarify that point.
- line 102: P_{lambda} is the "emitted spectra of the LEDs", not "spectral response". 
- line 119: change "basic functions" by "basis functions".
- line 123: change "Munssel" by "Munsell".
- In order to evaluate the spectral match between the original and the recovered one, the authors use the PSNR metric, which however only evaluates intensity differences. Other important correlation metrics should be used, such as GFC or SMAPE.

[1] M. Lopez, J. Hernandez, E. Valero, and J. Romero, “Selecting algorithms, sensors, and linear bases for optimum spectral recovery of skylight,” J. Opt. Soc. Am. A 24, 942–956 (2007).
[2] L. Arias, D. Sbarbaro, and S. Torres, “Removing baseline flame’s spectrum by using advanced recovering spectrum techniques,” Appl. Opt. 51, 6111–6116 (2012).
[3] C. Toro, L. Arias, S. Torres and D. Sbarbaro, “Flame spectra-temperature estimation based on a color imaging camera and a spectral reconstruction technique”, Applied Optics, 53(28), 6351-6361 (2014).

Author Response

We greatly appreciate the reviewer’s comments. We have made revisions accordingly.

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Dear,
although the scientific problem is not enough relevant due to indicated in my first revision (the algorithm improved by the authors is currently deprecated at all, and overpassed by other new algorithms), however, the new (corrected) manuscript provides more information concerning the problem. Also, the authors provide appropriate responses to all of my questions. Thus, I propose to accept the manuscript in its current form.