A Comparative Study of Filtering Approaches Applied to Color Archival Document Images

Maroua Mehri; Mohamed Ali Mahjoub; Walid Elhedda

REVIEW 3 major objections 6 minor 13 references

A Comparative Study of Filtering Approaches Applied to Color Archival Document Images

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that for blind denoising of color archival document images, marginal channel-by-channel filtering outperforms vector filtering, while for edge detection the vector approach is the better choice.

desk verdict A thin, flawed comparison of marginal vs. vector filtering: the vector baseline is scalar, the experiments are unreproducible, and the central claim is already standard knowledge. read the letter →

arxiv 1908.09007 v1 pith:5FUV3Z3J submitted 2019-08-16 cs.CV cs.LG

Walid Elhedda , Maroua Mehri , Mohamed Ali Mahjoub This is my paper · ORCID

classification cs.CVcs.LG

keywords colorimagefilteringmarginalapproachvectordocumentdenoisingedgedetectionarchivaldocumentshistoricalanalysismedianfilter

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper compares two ways of filtering color archival document images before analysis: filtering each color channel separately (the marginal approach) and filtering the pixel's color vector as a whole (the vector approach). Its experiments on noisy digitized archival pages show that marginal median and morphological filters produce higher PSNR and better region-statistics scores for denoising, while vector Laplacian and Sobel filters score better for edge detection. The paper concludes that for blind enhancement of historical documents, where the noise is unknown, marginal filtering is the safer default, and that vector filtering remains the right choice when the goal is to locate edges. The interest is a practical rule for choosing a preprocessing strategy without needing a noise model.

What carries the argument

The central object is the distinction between marginal and vector treatment of a color pixel. For order-statistic filters such as median and morphological erosion and dilation, the vector treatment sorts the norms of the RGB or HSB vectors, while the marginal treatment sorts each channel's intensities separately, and this difference drives the denoising results. For edge detection, the Laplacian and Sobel filters are evaluated through per-channel gradient terms combined under a square-root norm in the vector case, versus gradient-of-each-channel-then-combine in the marginal case. The RSC metric, computed from region statistics along four scan directions, is the criterion that separates the two approaches in the edge-detection experiments.

What would settle it

On a corpus of color archival document images with known text, add the six noise models, apply the best marginal denoising filter and the best vector one, then run an OCR system on both outputs. If the vector-filtered images produce equal or better word-level transcription despite lower PSNR and SR, the paper's metric-based ranking would be contradicted as a guide for the stated use case.

Watch

Extended reading notes

Core claim

The paper's central claim is a split recommendation: for denoising or enhancing color archival document images without a priori knowledge of the noise, apply the filter marginally, channel by channel; for edge detection, apply the filter vectorially, treating each pixel's color as one vector. The evidence comes from median and morphological filters under six noise types in RGB and HSB color spaces, with PSNR and the region-statistics ratio SR favoring the marginal approach for almost all noise types, and the RSC edge-statistics metric favoring the vector approach for Laplacian and Sobel filters. The paper also finds that dual sequences (marginal-then-vector or vector-then-marginal) do not improve denoising, and that morphological edge detection gives similar results under both approaches.

Load-bearing premise

The whole comparison rests on treating PSNR, SR, and RSC as sufficient proxies for filter quality; the paper never tests whether higher scores on these metrics actually improve OCR transcription, so the practical recommendation would fail if those metrics do not track the archival-document analysis goal.

Editorial extensions

If this is right

A preprocessing pipeline for noisy archival color documents can default to marginal median or morphological filtering, bypassing the need to estimate a noise model.
Edge-detection stages in the same pipeline should use vector Laplacian or Sobel filters, since combining channel gradients before computing magnitude preserves edges that marginal filtering splits across planes.
The color space, RGB or HSB, does not change the relative ranking, so the guidance is color-space independent.
Chaining marginal and vector passes in either order is not a worthwhile denoising strategy.
If a reliable noise model is available, vector filtering can outperform marginal filtering, so the blind condition is part of the recommendation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper never measures OCR accuracy, an implication left implicit is that the practical benefit should be verified end-to-end: the metric ranking may or may not translate into better transcription of degraded documents.
The same marginal-versus-vector logic plausibly extends to other nonlinear and order-statistic filters, but that is an extrapolation from the median and morphological evidence rather than a result of this paper.
The vector-norm sorting used in the vector median and morphological filters changes which pixel is selected, so it can alter color perception; a downstream-task or perceptual test would sharpen the guidance.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper's central claim—marginal filtering for denoising, vector for edge detection—is correct, but the paper's own evidence for the denoising half is invalid. Eq. 3 defines the 'vector' median as the median of vector norms, which is a scalar, not a vector. Same for morphological erosion/dilation in Eqs. 4-5. So the comparison pits a real marginal filter against a mislabeled scalar operator, and the PSNR in Eq. 14 assumes a vector output that Eq. 3 never produces. That's a load-bearing flaw, not a quibble. For the denoising conclusion, the paper's experiments don't actually compare what it says they compare.

What is genuinely new here is almost nothing. The conceptual categories (scalar, marginal, vector, hybrid) and the general guidance are decades old in color image processing. The only fresh element is applying these known filters to Tunisian archival documents, which is a reasonable application but not a new result. I'll give credit where it's due: the paper correctly notes that mean and Gaussian filters give identical results under marginal and vector approaches, and the edge-detection discussion (vector better because edges shouldn't be split across planes) is plausible. But those are observations, not contributions.

The experiments have other problems. Corpus size is never stated. Noise parameters are only labeled 'weak' or 'strong' with no quantitative definitions. No error bars, no statistical tests. The evaluation metrics (PSNR, SR, RSC) are never validated against OCR accuracy, which is the stated downstream goal. The figures are qualitative. For a paper whose entire purpose is to provide 'comprehensive guidelines,' this level of reporting is far too thin.

The self-citation [4] is about color space choice, different from this topic, so that's a minor issue. The larger issue is that the paper is not reproducible and, for the denoising claim, not internally consistent.

Who is this for? Maybe a practitioner who wants a rule of thumb and hasn't read the color filtering literature. But that reader would get the same guidance from a textbook. A serious referee would catch the Eq. 3 problem immediately, and the absence of experimental detail would sink it anyway.

My recommendation: desk reject. If the authors fixed the vector median definition, released code and data, quantified noise, and validated against OCR, this could become a minor workshop paper. As submitted, it doesn't warrant referee time.

Referee Report

3 major / 6 minor

Summary. This paper presents a comparative study of filtering approaches for color archival document images, focusing on marginal and vector implementations of denoising filters (median and morphological) and edge detection filters (Laplacian, Sobel, morphological). Experiments are carried out on images from the Tunisian national archives with six noise models, and performance is measured using PSNR, SR, and RSC metrics. The paper concludes that a marginal approach is best for denoising/enhancement, while a vector approach is better for edge detection, and that the choice should depend on whether a priori knowledge of image characteristics is available.

Significance. If the experimental comparison were sound, the paper would provide practical guidance for preprocessing color historical documents, a relevant problem for OCR and document image analysis. The paper's strengths include the use of multiple filters, two color spaces, and several noise types, as well as a quantitative evaluation rather than purely qualitative visual assessment. However, the central denoising claim rests on a mis-specified 'vector' filter, and the experimental protocol is insufficiently detailed to support the broad conclusions. The current version therefore does not yet establish the claimed practical guidance.

major comments (3)

[Section 2.1.2, Eq. (3) and Section 2.1.3, Eqs. (4)-(5)] The 'vector median filter' defined in Eq. (3) is not a vector filter: it computes the median of the norms of the neighboring color vectors, yielding a scalar output. A standard vector median filter selects the neighbor vector that minimizes the sum of distances to all other neighbors, producing a color vector. The same problem affects the morphological operations in Eqs. (4) and (5), where the minimum/maximum of vector norms is scalar-valued. This creates an internal inconsistency with the PSNR computation in Eq. (14), which sums per-channel squared differences and therefore assumes a vector-valued filtered image. Consequently, the comparisons in Figures 3-8 do not evaluate a genuine vector denoising approach, and the Section 5 conclusion that a marginal approach is best for denoising is not supported by the experiments as implemented.
[Section 4.1] The experimental protocol is not reproducible. The corpus size and composition are not stated, the noise parameters for 'weak' and 'strong' Gaussian, speckle, and salt-and-pepper noise are not quantified, and it is unclear whether the same images were used across all conditions. Figures 3-9 report metric values without error bars, confidence intervals, or statistical tests, yet the text repeatedly claims superiority 'for almost all types of noise.' The authors should disclose the full experimental setup and either provide variability measures or perform significance tests to support the claimed differences.
[Section 3 and Section 5] The paper motivates the work by the needs of OCR transcription for Tunisian archival documents, but it never measures OCR accuracy. The metrics PSNR, SR, and RSC are assumed to be valid proxies for downstream enhancement quality, but this assumption is not validated. Without evidence that improvements in these metrics translate to better OCR or human readability, the practical recommendation that a marginal approach is 'the best alternative for enhancing or denoising historical document images' is not established. The authors should either add an OCR evaluation or explicitly state and justify the assumed relevance of the chosen metrics.

minor comments (6)

[Section 5] The concluding statement that 'if we have a priori knowledge ... a vector approach will outperform a marginal one' is not supported by any experiment in the paper, since no condition varies the availability of a priori knowledge. This claim should be removed or qualified as speculation.
[Equation (1)] The notation in Eq. (1) is unclear: the arrows over 'OG' and 'OMi' are not defined, and the meaning of the centroid in the context of color image filtering should be explained more explicitly.
[Section 3.2] The definition of the SR metric is vague. Please specify how the regions are identified, how the standard deviation is computed per region, and whether the statistical mean m is computed per color channel or over the whole image.
[Section 3.3] The statement that the RSC metric 'uses the Lee filter to crop an image into regions' is confusing, as the Lee filter is a speckle reduction filter, not a region segmentation method. Please clarify the actual procedure used to define the statistical regions.
[Throughout] There are several typos and nomenclature issues: 'Laplacien' should be 'Laplacian', 'Prewit' should be 'Prewitt', 'Kirsh' should be 'Kirsch', and 'Derich' should be 'Deriche'. Also, in the conclusion, 'a filtering technique that do not require' should be 'a filtering technique that does not require'.
[Figure 9] The annotation description in Figure 9 is inconsistent: the text says the morphological, Sobel, and Laplacian filters are annotated with 'laplace', 'sobel', and 'morph', respectively, which is reversed relative to the filter order.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the comparative conclusions are empirically derived, not built into the filter definitions or metrics.

full rationale

This is an empirical benchmark study, not a derivation. The filters (median, morphological, mean, Laplacian, Sobel) are standard tools, and the evaluation metrics (PSNR, SR, RSC) are external criteria applied after filtering. The Section 5 conclusion that a marginal approach is best for blind denoising and a vector approach for edge detection is an inductive summary of the experimental results in Figures 3-9, not a quantity obtained by definition from the inputs. The only author self-citation ([4], Elhedda 2009) is background about color-space choice for segmentation and is not used to derive the main conclusion. The internal inconsistency that the 'vector' median and morphological filters in Eqs. 3-5 output scalar norms while the PSNR in Eq. 14 assumes vector-valued output is a methodological defect that could undermine the comparison, but it does not make the conclusion equivalent to its inputs by construction. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via self-citation. Consequently, no circular step is present.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities or fitted parameters. Its central claim rests on three domain assumptions: metric validity, noise model fidelity, and corpus representativeness. All are unvalidated and load-bearing for the practical conclusions.

assumptions (3)

domain assumption The evaluation metrics PSNR, SR, and RSC are valid indicators of filter performance for archival document image analysis.
Section 3 defines these metrics and uses them as the sole basis for judging filters, without validating that they correlate with OCR or transcription outcomes.
domain assumption Synthetic noise models (Gaussian, speckle, salt and pepper) adequately represent the degradations found in real historical documents.
Section 4.1 generates images by adding these noise types, but no validation is provided that these match aging, stains, bleed-through, or uneven lighting in archival scans.
domain assumption The corpus collected from the Tunisian national archives is representative of historical document images in general.
Section 4.1 states the corpus was collected from the Tunisian national archives but does not specify its size, composition, or representativeness, and the conclusions in Section 5 generalize to 'historical document images'.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Comparative Study of Filtering Approaches Applied to Color Archival Document Images." pith.science (2026). https://pith.science/paper/5FUV3Z3J

@misc{pith2026190809007,
  author       = {Pith},
  title        = {Pith review of: A Comparative Study of Filtering Approaches Applied to Color Archival Document Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FUV3Z3J}},
  note         = {Machine review of arXiv:1908.09007}
}

read the original abstract

Current systems used by the Tunisian national archives for the automatic transcription of archival documents are hindered by many issues related to the performance of the optical character recognition (OCR) tools. Indeed, using a classical OCR system to transcribe and index ancient Arabic documents is not a straightforward task due to the idiosyncrasies of this category of documents, such as noise and degradation. Thus, applying an enhancement method or a denoising technique remains an essential prerequisite step to ease the archival document image analysis task. The state-of-the-art methods addressing the use of degraded document image enhancement and denoising are mainly based on applying filters. The most common filtering techniques applied to color images in the literature may be categorized into four approaches: scalar, marginal, vector and hybrid. To provide a set of comprehensive guidelines on the strengths and weaknesses of these filtering approaches, a thorough comparative study is proposed in this article. Numerical experiments are carried out in this study on color archival document images to show and quantify the performance of each assessed filtering approach.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[4]

Color space choice for color images segmentation,

Elhedda, W., “Color space choice for color images segmentation,” in Colloque de la Recherche Appliquée et de Transfert de Technologie, 2009

work page 2009
[1]

Color segmentation by ordered mergings,

Angulo, J., and Serra, J., “Color segmentation by ordered mergings,” in International Conference on Image Processing, pp. 14-17, 2003

work page 2003
[2]

Restoring Ink Bleed -Through Degraded Document Images Using a Recursive Unsupervised Classification Technique,

Drira, F., Lebourgeois, F., and Emptoz, H., “Restoring Ink Bleed -Through Degraded Document Images Using a Recursive Unsupervised Classification Technique,” in International 8 Workshop on Document Analysis Systems, pp. 38- 49, 2006

work page 2006
[3]

Hermite and Gabor transforms for noise reduction and handwriting classification in ancient manuscripts,

Eglin, V., Bres, S., and Rivero, C., “Hermite and Gabor transforms for noise reduction and handwriting classification in ancient manuscripts,” International Journal of Document Analysis and Recognition, vol. 9, no. 2-4, pp. 101-122, 2007

work page 2007
[5]

Theory of vector filters based on linear quaternion functions,

Ell, T. A., and Sangwine, S. J., “Theory of vector filters based on linear quaternion functions,” in European Signal Processing Conference , pp. 1 - 5, 2008

work page 2008
[6]

M., Color in Computer Vision: Fundamentals and Applications

Gevers, T., Gijsenij, A., V andeweijer, J., and Geusebroek, J. M., Color in Computer Vision: Fundamentals and Applications . John Wiley & Sons, 2012

work page 2012
[7]

Restoration of ancient colored documents, foreground/background separation,

Ghomrassi, A., Charrada, M. A., and Amara, N. E. B., “Restoration of ancient colored documents, foreground/background separation,” in International Multi -Conference on Systems, Signals and Devices, pp. 1-6, 2015

work page 2015
[8]

Use of SLIC superpixels for ancient document image enhancement and segmentation,

Mehri, M., Sliti, N., Héroux, P., Gomez- Krämer, P., Amara, N. E. B., and Mullot, R., “Use of SLIC superpixels for ancient document image enhancement and segmentation,” in Document Recognition and Retrieval, vol. 9402, 2015

work page 2015

Show all 13 references

[9]

A review on the image enhancement of old historical documents and images,

Rana, B. , Kapoor, N., and Kundra, H., “A review on the image enhancement of old historical documents and images,” International Journal of Advances in Science and Technology, vol. 4, no. 1, pp. 180-185, 2016

2016
[10]

Colour image filters based on hypercomplex convolution,

Sangwine, S. J., and Ell, T. A., “Colour image filters based on hypercomplex convolution,” IEEE Proceedings - Vision, Image and Signal Processing, vol. 147, no. 2, pp. 89-93, 2000

2000
[11]

Restoration of old document images using different color spaces restoration of old document images,

Sgarbi, E. M., Mura, W. A. D., Moya, N. and Facon, J., and Ayala, H. A. L., “Restoration of old document images using different color spaces restoration of old document images,” in International Conference on Computer Vision Theory and Applications, pp. 82-88, 2014

2014
[12]

C olor Pixels Classification in an Hybrid Color Space,

Vandenbroucke, N., Macaire, L., and Postaire, J. G., “C olor Pixels Classification in an Hybrid Color Space,” in International Conference on Image Processing, pp. 176-180, 1998

1998
[13]

Review on image enhancement methods of old manuscript with damaged background,

Yahya, S. R., Abdullah, S. N. H. S., Omar, K., Zakaria, M. S. , Liong, C. Y. , “Review on image enhancement methods of old manuscript with damaged background,” in International Conference on Electrical Engineering and Informatics , pp. 62- 67, 2009. Walid Elhedda Walid Elhedda...

2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [4]

Color space choice for color images segmentation,

Elhedda, W., “Color space choice for color images segmentation,” in Colloque de la Recherche Appliquée et de Transfert de Technologie, 2009

work page 2009

[2] [1]

Color segmentation by ordered mergings,

Angulo, J., and Serra, J., “Color segmentation by ordered mergings,” in International Conference on Image Processing, pp. 14-17, 2003

work page 2003

[3] [2]

Restoring Ink Bleed -Through Degraded Document Images Using a Recursive Unsupervised Classification Technique,

Drira, F., Lebourgeois, F., and Emptoz, H., “Restoring Ink Bleed -Through Degraded Document Images Using a Recursive Unsupervised Classification Technique,” in International 8 Workshop on Document Analysis Systems, pp. 38- 49, 2006

work page 2006

[4] [3]

Hermite and Gabor transforms for noise reduction and handwriting classification in ancient manuscripts,

Eglin, V., Bres, S., and Rivero, C., “Hermite and Gabor transforms for noise reduction and handwriting classification in ancient manuscripts,” International Journal of Document Analysis and Recognition, vol. 9, no. 2-4, pp. 101-122, 2007

work page 2007

[5] [5]

Theory of vector filters based on linear quaternion functions,

Ell, T. A., and Sangwine, S. J., “Theory of vector filters based on linear quaternion functions,” in European Signal Processing Conference , pp. 1 - 5, 2008

work page 2008

[6] [6]

M., Color in Computer Vision: Fundamentals and Applications

Gevers, T., Gijsenij, A., V andeweijer, J., and Geusebroek, J. M., Color in Computer Vision: Fundamentals and Applications . John Wiley & Sons, 2012

work page 2012

[7] [7]

Restoration of ancient colored documents, foreground/background separation,

Ghomrassi, A., Charrada, M. A., and Amara, N. E. B., “Restoration of ancient colored documents, foreground/background separation,” in International Multi -Conference on Systems, Signals and Devices, pp. 1-6, 2015

work page 2015

[8] [8]

Use of SLIC superpixels for ancient document image enhancement and segmentation,

Mehri, M., Sliti, N., Héroux, P., Gomez- Krämer, P., Amara, N. E. B., and Mullot, R., “Use of SLIC superpixels for ancient document image enhancement and segmentation,” in Document Recognition and Retrieval, vol. 9402, 2015

work page 2015