{"id":"904da1f1-8600-4f33-aacb-8e95f2958405","arxiv_id":"1908.09007","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On color archival document images, marginal (per-channel) filtering outperforms vector approaches for denoising, while vector approaches outperform marginal for edge detection.","lead":"This paper compares four ways of applying standard denoising and edge-detection filters to color scans of historical Arabic documents, and finds that processing each color channel separately works best for reducing noise while vector-based processing works best for finding edges. It is a practical benchmarking study for the Tunisian national archives' OCR pipeline, but it shares no code, data, or error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'vector' denoising baseline is not a vector filter: Eq. 3 returns a scalar median-of-norms, so the central marginal-vs-vector denoising claim rests on an invalid comparison.","rationale":"The reader's weakest_assumption concerns the validity of PSNR/SR/RSC as proxies for OCR quality; that is a real external-validity concern. However, the more load-bearing issue is internal: the paper's vector denoising filters, as defined in Eqs. 3-5, are not vector filters at all. The central claim about marginal superiority for denoising is drawn from experiments in which the vector baseline is a scalar norm-based operator, so the comparison does not test what the conclusion claims. This is a correctness risk in the core evidence, not merely a missing downstream validation. I still do not think the paper must be rejected outright: the authors could correct the vector filter definitions, rerun the experiments, and possibly reach the same conclusion. That is why the verdict should remain CONDITIONAL rather than move to REJECT. The reader did mention the incorrect vector median definition in the rationale, but did not make it the weakest_assumption, hence partial agreement. The proposed concrete test would settle whether the central claim survives a corrected vector implementation.","tokens_in":9735,"tokens_out":3586,"duration_ms":37798,"concrete_test":"Re-implement the median and morphological filters under a genuine vector approach: for each 3x3 neighborhood, define the vector median as the input vector minimizing the sum of L2 distances to the other eight vectors (or, for morphological erosion/dilation, select the vector whose norm is minimum/maximum while retaining its full color components). Rerun the PSNR and SR experiments for Noise 1-6 in RGB and HSB. If the marginal approach does not remain best for the majority of noise/filter/metric combinations, the Section 5 denoising claim is refuted; if it does remain best, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim in Section 5 — that a marginal approach is the best alternative for denoising/enhancing historical document images — is directly supported by Figures 3-8, which compare marginal, vector, and dual implementations of the median and morphological filters. But the 'vector' implementations are not vector filters. Section 2.1.2 defines the vector median filter as F(x,y) = median(||a||, ..., ||i||) (Eq. 3), and Section 2.1.3 defines vector morphological erosion/dilation as the minimum/maximum of vector norms (Eqs. 4-5). These operations output a single scalar norm value, not a color vector. They are scalar operators applied to a norm image, so the comparison pits a proper marginal filter against a mislabeled scalar operator. The inconsistency is internal: the PSNR definition in Eq. 14 sums per-channel squared differences, which assumes a vector-valued filtered image, whereas Eq. 3 yields a scalar. This matters because the paper's own motivation for vector filtering is to preserve the multidimensional color information, and a standard vector median (e.g., reduced ordering selecting the pixel vector with minimum aggregate distance to its neighbors) can behave differently on exactly the salt-and-pepper and speckle noise cases tested. If the vector baseline had been implemented correctly, the observed ranking that drives the Section 5 conclusion could change. This is not a disagreement with consensus; it is a defect in the paper's own experimental construction.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9960,"tokens_out":4570,"duration_ms":41994,"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":[{"comment":"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":"Section 2.1.2, Eq. (3) and Section 2.1.3, Eqs. (4)-(5)"},{"comment":"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":"Section 4.1"},{"comment":"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.","section":"Section 3 and Section 5"}],"minor_comments":[{"comment":"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.","section":"Section 5"},{"comment":"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":"Equation (1)"},{"comment":"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":"Section 3.2"},{"comment":"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.","section":"Section 3.3"},{"comment":"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'.","section":"Throughout"},{"comment":"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.","section":"Figure 9"}],"recommendation":"major_revision","confidential_remarks":"The vector median filter issue is fundamental and would be immediately apparent to a reviewer familiar with color image filtering. The authors should consult standard references on vector median filters (e.g., Astola et al., 1990) and rerun the experiments with a correct implementation. Additionally, the paper's claim to provide 'comprehensive guidelines' is stronger than the evidence supports given the narrow set of filters and the lack of experimental details."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat 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.\n\nThe 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.\n\nThe 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.\n\nWho 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.\n\nMy 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.","headline":"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.","tokens_in":10512,"tokens_out":1456,"would_cite":false,"duration_ms":16327,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["color image filtering","marginal approach","vector approach","document image denoising","edge detection","archival documents","historical document image analysis","median filter"],"falsifier":"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.","tokens_in":9496,"feed_emoji":"📜","tokens_out":6011,"duration_ms":55912,"temperature":0.7,"pith_summary":"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.","feed_headline":"Marginal filters beat vector filters for denoising old color documents","feed_subtitle":"For finding edges, though, the full color vector wins. A concrete guide for preprocessing archival scans.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"supplies the multidimensional color morphology formalism that underlies the vector morphological filter tested here.","marker":"[1]"},{"why":"introduces vector filters for color images based on quaternion functions, establishing the vector-approach family being compared.","marker":"[5]"},{"why":"introduces hypercomplex color convolution and motivates the paper's claim that such vector-style color filtering requires a priori knowledge.","marker":"[10]"},{"why":"provides the RSC region-statistics metric and the hybrid-color-space classification ideas used to evaluate edge-detection filters.","marker":"[12]"},{"why":"treats restoration of degraded color document images as a distinct problem, supporting why color-aware filtering is needed.","marker":"[2]"}],"fun_headline_variants":["For old docs: marginal filters denoise, vector filters detect edges","Split filtering: marginal for noise, vector for edges in scans","Marginal beats vector for denoising, but vector wins for edges","Best filter depends on task: marginal denoise, vector edge detect"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["For old docs: marginal filters denoise, vector filters detect edges","Split filtering: marginal for noise, vector for edges in scans","Marginal beats vector for denoising, but vector wins for edges","Best filter depends on task: marginal denoise, vector edge detect"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1630,"prompt_tokens":874,"completion_tokens":756,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":681}},"tokens_in":490,"tokens_out":756,"duration_ms":6437,"temperature":1.0,"reasoning_tokens":681,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:59:06.799767+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Color segmentation by ordered mergings,","cited_arxiv_id":null,"evidence_quote":"supplies the multidimensional color morphology formalism that underlies the vector morphological filter tested here."},{"cited_title":"Theory of vector filters based on linear quaternion functions,","cited_arxiv_id":null,"evidence_quote":"introduces vector filters for color images based on quaternion functions, establishing the vector-approach family being compared."},{"cited_title":"Colour image filters based on hypercomplex convolution,","cited_arxiv_id":null,"evidence_quote":"introduces hypercomplex color convolution and motivates the paper's claim that such vector-style color filtering requires a priori knowledge."},{"cited_title":"C olor Pixels Classification in an Hybrid Color Space,","cited_arxiv_id":null,"evidence_quote":"provides the RSC region-statistics metric and the hybrid-color-space classification ideas used to evaluate edge-detection filters."},{"cited_title":"Restoring Ink Bleed -Through Degraded Document Images Using a Recursive Unsupervised Classification Technique,","cited_arxiv_id":null,"evidence_quote":"treats restoration of degraded color document images as a distinct problem, supporting why color-aware filtering is needed."}],"review_version":1}