A Method for Paired Comparisons of Glo Germ Quantity in Images of Hands Before and After Washing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a reproducible pipeline that converts color images into quantitative fluorescence maps by combining spectral measurement with a linear mixture model. Taking Glo-Germ as the research object, it measures the emission spectrum with a spectral photometer and normalizes it, projects the spectrum into CIE XYZ coordinates and incorporates it into the linear mixture model, and estimates the contributions of white light, UV-illuminated skin reflectance and fluorophore emission in each pixel using the non-negative least squares method, yielding a grayscale image that can characterize the local fluorophore density. It also obtains an image-level summary result proportional to the total amount of detected substances through spatial integration. Compared with single-channel proxy methods, this method can suppress background structures, improve contrast and maintain radiometric interpretability. Relying only on measurable spectra and linear transforms, the method can be reproduced across different cameras and extended to other fluorophores.
(1) The study fails to clearly explain the essential differences between the core research questions it explores and those of existing fluorescence detection studies, nor does it adequately demonstrate the originality of the research topic in the field of fluorescence imaging. It is necessary to specifically illustrate the research gaps addressed by this study and its differentiated value compared with existing research.
(2) The implementation details of the reproducibility research for the proposed method are insufficient. It does not elaborate on the specific operational steps and calibration standards for converting data from cameras of different brands and models into the CIE XYZ color space, nor does it provide access to the model code, original spectral data and image datasets, which hinders reproducibility and verification by other researchers.
(3) The evaluation and validation research of the method has limitations. Only taking Glo-Germ as the research object, the method is not validated with various fluorophores, different surface materials and under complex illumination conditions, nor is it quantitatively compared with well-known baseline methods in this field (such as the traditional color threshold method and the detection method using professional fluorescence imagers), making it impossible to fully prove the universality and superiority of the method.
(4) Further reference can be made to the learning method of robust feature representation in "From simple to complex scenes: Learning robust feature representations for accurate human parsing", which provides ideas for feature extraction in complex backgrounds for fluorescence detection and improves the accuracy of distinguishing skin reflection from fluorescence signals. Meanwhile, the anti-geometric attack technology in "Light-Field Image Multiple Reversible Robust Watermarking Against Geometric Attacks" can be used for reference to optimize the signal processing process of fluorescence images, enhance the detection robustness of the model under different illumination conditions and shooting angles, and improve the practical application performance of the method.

Author Response

1.  The study fails to clearly explain the essential differences between the core research questions it explores and those of existing fluorescence detection studies, nor does it adequately demonstrate the originality of the research topic in the field of fluorescence imaging. It is necessary to specifically illustrate the research gaps addressed by this study and its differentiated value compared with existing research.
   1. We have narrowed the scope of the paper to specifically dealing with paired comparisons of the same hands before and after washing, with the intention of measuring specifically GloGerm on specifically human hands under specific UV illumination.  All of the necessary parameters for taking the images are discussed.  
2. The implementation details of the reproducibility research for the proposed method are insufficient. It does not elaborate on the specific operational steps and calibration standards for converting data from cameras of different brands and models into the CIE XYZ color space, nor does it provide access to the model code, original spectral data and image datasets, which hinders reproducibility and verification by other researchers.
   1. In addition to the narrowed scope, we have included more information about our particular camera, and we have included code in the supplementary materials, including a MATLAB function that only requires a RAW file name as input, so readers can see exactly how we obtain the XYZ image.  We have also included supplementary materials of the processed spectral measurements and image datasets.  
3. The evaluation and validation research of the method has limitations. Only taking Glo-Germ as the research object, the method is not validated with various fluorophores, different surface materials and under complex illumination conditions, nor is it quantitatively compared with well-known baseline methods in this field (such as the traditional color threshold method and the detection method using professional fluorescence imagers), making it impossible to fully prove the universality and superiority of the method.
   1. By narrowing the scope, hopefully taking GloGerm as the only research object is not a problem.  We are specifically interested in quantifying the success of a handwashing training regime by measuring the amount of GloGerm removed from the skin by washing.     We have included for comparison the matched filter response, which we believe is a reliable baseline for inferring how other methods would perform.  The problem with thresholding is that in converts the images into a binary format, and thus loses information that discriminates between above-threshold pixels.  Unfortunately, we do not have access to professional imaging technology, so we cannot compare. 
4.  Further reference can be made to the learning method of robust feature representation in "From simple to complex scenes: Learning robust feature representations for accurate human parsing", which provides ideas for feature extraction in complex backgrounds for fluorescence detection and improves the accuracy of distinguishing skin reflection from fluorescence signals. Meanwhile, the anti-geometric attack technology in "Light-Field Image Multiple Reversible Robust Watermarking Against Geometric Attacks" can be used for reference to optimize the signal processing process of fluorescence images, enhance the detection robustness of the model under different illumination conditions and shooting angles, and improve the practical application performance of the method.
   1. Liu et al., (2024) appears to focus on the segmentation of human body parts in computer vision.  This kind of segmentation problem necessarily includes between-pixel information (i.e. an edge only emerges from a sequence of pixels), while our focus is on "segmenting" the light captured by a single pixel.   While our focus is not segmentation, Liu et al., provide a useful tool that could be used to crop hands from the background in our data sets, so we included the citation where appropriate in the discussion.  
   2. Wang et al., (2025) appears to focus on ownership verification of images through robust water-marking.  It seems to me that these authors are concerned with the detection of a binary signal (watermark or no watermark) in a whole image, whereas we are concerned with the detection of a continuous signal at the single pixel level.  This paper addresses robustness and reversibility in signal embedding using transform-domain invariants and redundancy, whereas the revised manuscript addresses physical signal separation using a forward model of image formation.  We could not find a place for this citation in the paper.  

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes a reproducible pipeline that converts color images into quantitative fluorescence maps by combining spectral measurement with a linear mixture model.

The weaknesses of this paper are as follows:

1. The introduction lacks clarity in explaining the motivation for combining spectral measurement with a linear mixture model and does not list the main contributions in a structured manner.
2. The related work section is not clearly defined; the authors should give a separate section to outline existing techniques and methods and analyze their limitations.
3. The experimental results section only shows a few figures (e.g., Figures 7-14), which seems insufficient to provide strong proof. Simply demonstrating that some color images can be quantitatively converted into fluorescence maps only shows feasibility and does not represent the robustness and effectiveness of the method. It is recommended to add statistical experimental analysis (conducting experiments on a large image dataset) to present corresponding results.
4. Additionally, regarding the authors' proposed method for converting color images into fluorescence images, it is unclear whether consideration has been given to the fact that images captured in real-world scenarios are often not clear. In such challenging cases, the authors should at least account for these extreme conditions by incorporating and reviewing recent image enhancement and restoration methods, such as IHDCP, RLP, etc to improve image quality before performing the conversion to fluorescence images

[1] IHDCP: Single Image Dehazing Using Inverted Haze Density Correction Prior

[2] A Dual-Stage Residual Diffusion Model With Perceptual Decoding for Remote Sensing Image Dehazing

5. The authors only presented images of palms; it is recommended to provide conversion results for a wider variety of different scenes.
6. Why not consider converting to other color spaces, such as HSV or YCbCr? It is recommended to provide a discussion and experimental comparison.
7. The biggest issue with the paper is that the experiments are insufficient, with no comparison to any prior work. The authors should add comparisons with other related works by providing comparative result figures and experiments.
8. There are some issues with the sentence structure and informal expressions in the paper. Additionally, it is recommended to include more recent references.

Author Response

1. The introduction lacks clarity in explaining the motivation for combining spectral measurement with a linear mixture model and does not list the main contributions in a structured manner.
   1. We have  added a section to the introduction titled "Contributions."  Additionally, we have clarified that the motivation for combining spectral measurement with a linear mixture model is the additivity of light caused by fluorescent emissions. 
2. The related work section is not clearly defined; the authors should give a separate section to outline existing techniques and methods and analyze their limitations.
   1. We have narrowed the scope of the paper to focus exclusively on images of human hands under specific illumination conditions with the goal of detecting GloGerm in order to determine the effectiveness of a particular handwashing routine.  
3. The experimental results section only shows a few figures (e.g., Figures 7-14), which seems insufficient to provide strong proof. Simply demonstrating that some color images can be quantitatively converted into fluorescence maps only shows feasibility and does not represent the robustness and effectiveness of the method. It is recommended to add statistical experimental analysis (conducting experiments on a large image dataset) to present corresponding results.
   1. We have added two sections to the results.  The first includes response curves taken at different exposure durations.  These curves are generated from test strip photos of variable concentration swatches, and these are included in the supplementary materials.  The second includes a mean-squared error analysis from simulated light spectra.  
4. Additionally, regarding the authors' proposed method for converting color images into fluorescence images, it is unclear whether consideration has been given to the fact that images captured in real-world scenarios are often not clear. In such challenging cases, the authors should at least account for these extreme conditions by incorporating and reviewing recent image enhancement and restoration methods, such as IHDCP, RLP, etc to improve image quality before performing the conversion to fluorescence images
   1. It is true that in the real world, images are often unclear.  However, we have narrowed the scope of our paper to focus on very specific conditions-- where the subject of the photo is always human hands with or without GloGerm, and the illuminant matches the illuminant used by us.  
   2. IHDCP and RLP are an interesting topic that is similar to our paper at the modeling level, although dehazing is a fundamentally different process than the mechanism we used to derive our model.  However, both approaches are based on a physical model of the process.  In IHDCP, that model is based on atmospheric scattering.  In our paper, it is based on the additivity of lights.  However, under our viewing conditions, there isn't enough distance or particulate matter suspended in the atmosphere to introduce hazing.  All of the images are taken from inside a black-out box where the subject of the photo is only a few inches away.  We have added a paragraph to the discussion to compare RLP and IHDCP to our method.  
5. The authors only presented images of palms; it is recommended to provide conversion results for a wider variety of different scenes.
   1. We have narrowed the scope of the paper to focus specifically on one type of scene.  The images provided in the paper include background features in an attempt to illustrate conversion results for a "wider" variety of scenes that could be observed under our imaging setup.  The controlled illuminant is critical here... it is a very narrowband light, which means that the photons available from the scene are limited, and thus when we consider "different" scenes, we have to rule out all scenes with a different illuminant.    In addition, we have included more photos in the supplementary material. 
6. Why not consider converting to other color spaces, such as HSV or YCbCr? It is recommended to provide a discussion and experimental comparison.
   1. The XYZ color-matching functions we used are "physiologically-relevant" in the sense that they are derived from a linear combination of cone fundamentals. Our main author is a psychophysicists and vision scientist, so he has lots of experience working with these functions.   They are the most commonly used colorimetric coordinates, with most of the alternative colorspaces being linear transformations of this basis.  HSV is an exception, with the nonlinearity involved and the resulting hue measure being represented in polar coordinates.    This use of a periodic variable makes linear regression problematic, since the periodic variable is the same at 0 degrees and 360 degrees, but a regression model would output very different values for the two.   Any other color space that results from a linear transformation of XYZ would simply impose a change of basis in our equations.  
7. The biggest issue with the paper is that the experiments are insufficient, with no comparison to any prior work. The authors should add comparisons with other related works by providing comparative result figures and experiments.
   1. We have included the matched filter response in our methods for comparison, and completed two more experiments.  The first experiment allows us to measure response curves using real images of swatches varying in concentration, and these images are included in the supplementary materials.  The second experiment is based on simulated light spectra and allows us to measure the mean-squared error of each method acting as an estimator for fluorescence.  
8. There are some issues with the sentence structure and informal expressions in the paper. Additionally, it is recommended to include more recent references.
   1. Hopefully we have improved the sentence structure, and we have added more recent references.  

Reviewer 3 Report

Comments and Suggestions for Authors

Title: A Spectrally Informed Linear Model for Fluorescence Detection in Color Images

The manuscript presents a solid, physically grounded framework for quantifying Glo Germ fluorescence using a linear mixture model in CIE XYZ space. Moving from heuristic thresholding to a spectral "model observer" is a significant step forward for hygiene and biosecurity research. The methodology is clear, and the use of commodity hardware makes it highly practical.

Major Points:

Theoretical Positioning: Please more explicitly contrast this method with classical Matched Filtering and Color Deconvolution. A small comparison table would help clarify how additive fluorescence modeling differs from subtractive staining approaches.

Quantitative Validation: While visual results are strong, the claim that beta_f is a linear proxy for density needs more empirical backing. Including a density-response curve (measured vs. predicted density) and characterizing the "saturation regime" mentioned in Appendix B would greatly strengthen the paper.

Sensitivity Analysis: Since the model relies on RAW-to-XYZ transforms, please add a brief discussion or simulation on how inaccuracies in the color transformation matrix or camera spectral sensitivities affect the beta_f estimates.

Model Robustness: The current model focuses on skin/white-card backgrounds. Briefly discuss how the detector performs on other common materials (e.g., nitrile gloves or fabrics) where the background spectrum x differs.

UV Leakage: Clarify if the 390–830 nm range consistently accounts for UV leakage below 400 nm.

Hardware Details: Move camera specifications from the appendix/subtext to the main Materials and Methods section.

Terminology: Rephrase "novel applications of the Beer-Lambert law" to "novel radiometric application" to avoid overstatement.

Data: Sharing the measured SPDF curves as supplementary material is highly recommended for reproducibility.

This is a high-quality study that bridges color science and practical imaging. Addressing the points above regarding quantitative validation and theoretical context will make it a very strong contribution to the field.

Author Response

1. Theoretical Positioning: Please more explicitly contrast this method with classical Matched Filtering and Color Deconvolution. A small comparison table would help clarify how additive fluorescence modeling differs from subtractive staining approaches.
   1. We have added the matched filter response to the paper for comparisons to our non-negative least squares approach.  Unfortunately, it is not possible to directly compare color deconvolution to our process because the physics of stain histology are subtractive (as opposed to additive), and depend on an absorbance parameter that is calculated from the ratio of intensity entering a sample to intensity exiting the sample.  There is no analogous measurement setup in our situation, and no analogous parameter to absorbance.
   2. We have added the table you requested. 
2. Quantitative Validation: While visual results are strong, the claim that beta_f is a linear proxy for density needs more empirical backing. Including a density-response curve (measured vs. predicted density) and characterizing the "saturation regime" mentioned in Appendix B would greatly strengthen the paper.
   1. We have added the density-response curves to the results section. 
3. Sensitivity Analysis: Since the model relies on RAW-to-XYZ transforms, please add a brief discussion or simulation on how inaccuracies in the color transformation matrix or camera spectral sensitivities affect the beta_f estimates.
   1. We have added a simulation that allows us to calculate the mean-squared error of the beta_f as a estimator for fluorescence in simulated lights.  By adding noise to the resulting colors, we can see how color components outside the space spanned by the model contribute to performance. 
4. Model Robustness: The current model focuses on skin/white-card backgrounds. Briefly discuss how the detector performs on other common materials (e.g., nitrile gloves or fabrics) where the background spectrum x differs.
   1. We have narrowed the scope of the paper to be focused on detecting GloGerm in images of human hands under controlled illumination conditions.  We have added some details about controlling these viewing conditions so that no alternative sources of light exist.  
5. UV Leakage: Clarify if the 390–830 nm range consistently accounts for UV leakage below 400 nm.
   1. I'm not sure what UV leakage means, but the 390-830 nm range is standard for tables of color matching functions.  Outside of this range, the functions usually take value of zero, so extending the range of light we actually measure would not change the tristimulus values we observe.  
6. Hardware Details: Move camera specifications from the appendix/subtext to the main Materials and Methods section.
   1. Done!
7. Terminology: Rephrase "novel applications of the Beer-Lambert law" to "novel radiometric application" to avoid overstatement.
   1. Done!
8. Data: Sharing the measured SPDF curves as supplementary material is highly recommended for reproducibility.
   1. Done! And much more included in the supplementary materials.  

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have carefully revised the manuscript according to my comments. I recommend that the paper be accepted for publication.

Reviewer 2 Report

Comments and Suggestions for Authors

All my concerns have been addressed. I recommend accept for this paper.