pith. sign in

arxiv: 2604.19510 · v1 · submitted 2026-04-21 · 💻 cs.CV

Evaluating Histogram Matching for Robust Deep learning-Based Grapevine Disease Detection

Ruben Pascual , In\'es Hern\'andez , Salvador Guti\'errez , Javier Tardaguila , Pedro Melo-Pinto , Daniel Paternain , Mikel Galar This is my paper

Pith reviewed 2026-05-10 02:28 UTC · model grok-4.3

classification 💻 cs.CV
keywords histogram matchingdeep learninggrapevine disease detectionillumination robustnessdata augmentationpreprocessingcanopy imagesplant pathology
0
0 comments X

The pith

Histogram matching used as both preprocessing and augmentation improves deep learning robustness for grapevine disease detection on real canopy images.

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

The paper tests histogram matching as a way to handle variable illumination, the main obstacle to reliable deep learning performance when classifying grapevine leaves in the field. It applies the technique twice: once to normalize all images to a single reference intensity profile, and again during training to create controlled lighting variations. Experiments with 1,469 images and ResNet-18 models show clear gains on heterogeneous canopy shots that include background and mixed lighting, while gains on clean leaf-focused shots remain small. A reader would care because this points to a lightweight way to make automated disease scouting practical without special lighting rigs or massive new datasets.

Core claim

Integrating histogram matching in two stages—first as a fixed normalization step that aligns each image's pixel intensities to a reference profile, and second as a data-augmentation operator that deliberately varies the profile during training—produces models whose accuracy on uncontrolled canopy images rises markedly compared with standard training, while the same models show only modest change on uniform leaf images.

What carries the argument

Histogram matching, which remaps the intensity values of an image so its distribution matches a chosen reference histogram; it is used both to standardize inputs and to diversify training examples.

If this is right

  • Field-deployed classifiers for grapevine diseases can maintain performance across different times of day and weather without requiring controlled illumination.
  • The largest robustness gains appear on complex canopy views rather than isolated leaves, suggesting the method targets the exact domain shift present in real scouting.
  • The same normalization-plus-augmentation pattern can be added to existing deep-learning pipelines for other crops facing lighting variability.
  • Training data requirements may decrease because the augmentation step synthetically supplies lighting diversity that would otherwise need to be collected in the field.

Where Pith is reading between the lines

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

  • The approach could be tested on other plant-disease datasets to check whether the canopy-specific benefit holds for different crop architectures.
  • Choosing the reference histogram from images captured at a representative time of day might further increase the method's effectiveness across seasons.
  • Combining histogram matching with other low-cost preprocessing steps, such as color constancy, might compound the robustness gains.

Load-bearing premise

The reference histogram chosen for matching represents the actual lighting conditions encountered in target vineyard images, and any observed accuracy gains come from the matching steps themselves rather than from other details of the training pipeline.

What would settle it

A follow-up experiment that applies the identical dual-stage histogram matching pipeline to a fresh collection of canopy images and finds accuracy no higher, or lower, than the baseline model trained without matching.

Figures

Figures reproduced from arXiv: 2604.19510 by Daniel Paternain, In\'es Hern\'andez, Javier Tardaguila, Mikel Galar, Pedro Melo-Pinto, Ruben Pascual, Salvador Guti\'errez.

Figure 1
Figure 1. Figure 1: Illustration of the proposed histogram matching strategies. (a) Histogram matching pre-processing: the cumulative distribution functions (CDFs) of all RGB channels are averaged across the dataset to obtain a reference mean CDF. Each image is then adjusted to match this reference. (b) Histogram matching data augmentation: during training, a random image from the original dataset is selected as reference, an… view at source ↗
Figure 2
Figure 2. Figure 2: Examples of leaf images from each image type [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of images before and after histogram [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

Variability in illumination is a primary factor limiting deep learning robustness for field-based plant disease detection. This study evaluates Histogram Matching (HM), a technique that transforms the pixel intensity distribution of an image to match a reference profile, to mitigate this in grapevine classification, distinguishing among healthy leaves, downy mildew, and spider mite damage. We propose a dual-stage integration of HM: (i) as a preprocessing step for normalization, and (ii) as a data augmentation technique to introduce controlled training variability. Experiments using 1,469 RGB images (comprising homogeneous leaf-focused and heterogeneous canopy samples) to train ResNet-18 models demonstrate that this combination significantly enhances robustness on real-world canopy images. While leaf-focused samples showed marginal gains, the canopy subset improved markedly, indicating that balancing normalization with histogram-based diversification effectively bridges the domain gap caused by uncontrolled lighting.

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

3 major / 1 minor

Summary. The manuscript evaluates histogram matching (HM) as a dual-stage technique—both as preprocessing normalization and as data augmentation—to mitigate illumination variability in deep learning-based classification of grapevine diseases (healthy leaves, downy mildew, spider mite damage). Experiments train ResNet-18 models on a total of 1,469 RGB images containing both homogeneous leaf-focused and heterogeneous canopy samples, with the central claim that the combination yields marked robustness gains on real-world canopy images while producing only marginal improvements on leaf-focused samples.

Significance. If the quantitative claims can be substantiated, the approach would represent a simple, computationally lightweight method for improving domain robustness in field-based agricultural vision systems, where uncontrolled lighting is a persistent challenge. The emphasis on canopy-level heterogeneity is practically relevant. However, the absence of supporting metrics, controls, and methodological specifics currently limits the work's contribution to the literature on robust plant disease detection.

major comments (3)
  1. [Abstract] Abstract: the claim that dual-stage HM 'significantly enhances robustness on real-world canopy images' and produces 'marked' improvement is unsupported by any quantitative results (accuracy/F1 deltas, error bars, p-values, or baseline comparisons without HM).
  2. [Methods] Methods/Experiments: no description is given of how the reference histogram profile is constructed or selected (source images, criteria, or statistics), leaving the representativeness assumption for target field conditions unverified and the method non-reproducible.
  3. [Experiments] Experiments: the manuscript provides no ablation that removes HM while holding data splits, other augmentations, optimizer settings, and model architecture fixed, so the causal link between the dual-stage HM and the reported canopy gains cannot be established.
minor comments (1)
  1. [Abstract] Abstract: the total of 1,469 images is stated but the breakdown between leaf-focused and canopy subsets, as well as the train/test partitioning strategy, is not specified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. The comments highlight important areas where the manuscript can be strengthened for clarity, reproducibility, and rigor. We address each major comment point by point below, indicating the specific revisions we will make.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that dual-stage HM 'significantly enhances robustness on real-world canopy images' and produces 'marked' improvement is unsupported by any quantitative results (accuracy/F1 deltas, error bars, p-values, or baseline comparisons without HM).

    Authors: We agree that the abstract would be strengthened by explicit quantitative support. In the revised manuscript, we will update the abstract to report the specific accuracy and F1-score improvements observed on the canopy subset (with and without dual-stage HM), including any available deltas, error bars, and baseline comparisons. This will directly substantiate the claims of significant enhancement and marked improvement. revision: yes

  2. Referee: [Methods] Methods/Experiments: no description is given of how the reference histogram profile is constructed or selected (source images, criteria, or statistics), leaving the representativeness assumption for target field conditions unverified and the method non-reproducible.

    Authors: We acknowledge that additional methodological detail is required for reproducibility. We will expand the Methods section to fully describe the construction of the reference histogram profile, specifying the source images selected (a representative subset of well-illuminated canopy images from the dataset), the selection criteria (e.g., median intensity distribution across channels), and the exact statistics used (e.g., cumulative distribution function matching). This will allow verification of the profile's suitability for target field conditions. revision: yes

  3. Referee: [Experiments] Experiments: the manuscript provides no ablation that removes HM while holding data splits, other augmentations, optimizer settings, and model architecture fixed, so the causal link between the dual-stage HM and the reported canopy gains cannot be established.

    Authors: We agree that an explicit ablation is necessary to isolate the contribution of dual-stage HM and establish the causal link. In the revised manuscript, we will add a controlled ablation study in which ResNet-18 models are trained with identical data splits, augmentations (excluding HM), optimizer settings, and architecture, comparing performance with and without the dual-stage HM. Results will be reported specifically for the heterogeneous canopy images to quantify the robustness gains attributable to HM. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical evaluation of standard HM technique

full rationale

The paper is an empirical evaluation study that applies the well-known histogram matching technique (both as preprocessing normalization and as data augmentation) to train a standard ResNet-18 classifier on a fixed set of 1,469 grapevine images. No mathematical derivation, fitted parameter, or first-principles result is presented that reduces to its own inputs by construction. Reported gains on canopy versus leaf subsets are experimental outcomes, not predictions forced by self-definition or self-citation chains. The central claim rests on standard supervised learning and direct comparison, with no equations or uniqueness theorems invoked.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that histogram matching can meaningfully reduce illumination-induced domain shift for CNN classification; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Histogram matching can effectively reduce illumination variability in RGB images for DL classification.
    Core premise enabling both preprocessing and augmentation stages.

pith-pipeline@v0.9.0 · 5470 in / 1142 out tokens · 31988 ms · 2026-05-10T02:28:06.350620+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

31 extracted references · 31 canonical work pages

  1. [1]

    State of the world vine and wine sector 2024,

    International Organisation of Vine and Wine (OIV), “State of the world vine and wine sector 2024,” 2025

    work page 2024

  2. [2]

    Galet,Grape Diseases

    P. Galet,Grape Diseases. Oeno Plurimedia, 1996

    work page 1996

  3. [3]

    W. F. Wilcox, W. D. Gubler, and J. K. Uyemoto,Compendium of Grape Diseases, Disorders, and Pests, Second Edition. The American Phytopathological Society, 2015

    work page 2015

  4. [4]

    The Course of Colonization of Two Different Vitis Genotypes by Plasmopara viticola Indicates Compatible and Incompatible Host-Pathogen Interactions,

    S. Unger, C. B ¨uche, S. Boso, and H.-H. Kassemeyer, “The Course of Colonization of Two Different Vitis Genotypes by Plasmopara viticola Indicates Compatible and Incompatible Host-Pathogen Interactions,”Phytopathology, vol. 97, no. 7, pp. 780–786, 2007

    work page 2007

  5. [5]

    Deep learning for the differentiation of downy mildew and spider mite in grapevine under field conditions,

    S. Guti ´errez, I. Hern ´andez, S. Ceballos, I. Barrio, A. M. D ´ıez- Navajas, and J. Tardaguila, “Deep learning for the differentiation of downy mildew and spider mite in grapevine under field conditions,”Comput. Electron. Agr ., vol. 182, p. 105991, 2021

    work page 2021

  6. [6]

    SVM-Based Detection of Tomato Leaves Diseases,

    U. Mokhtar, N. El Bendary, A. E. Hassenian, E. Emary, M. A. Mahmoud, H. Hefny, and M. F. Tolba, “SVM-Based Detection of Tomato Leaves Diseases,” inIntell. Syst.Cham: Springer, 2015, pp. 641–652

    work page 2015

  7. [7]

    From visual estimates to fully automated sensor-based measurements of plant disease severity: Status and challenges for improving accuracy,

    C. H. Bock, J. G. A. Barbedo, E. M. Del Ponte, D. Bohnenkamp, and A.-K. Mahlein, “From visual estimates to fully automated sensor-based measurements of plant disease severity: Status and challenges for improving accuracy,”Phytopathol. Res., vol. 2, no. 1, p. 9, 2020

    work page 2020

  8. [8]

    Development of rapid direct PCR assays to identify downy and powdery mildew and grey mould in vitis vinifera tissues,

    K. Gindroet al., “Development of rapid direct PCR assays to identify downy and powdery mildew and grey mould in vitis vinifera tissues,”J. Int. Sci. Vigne Vin, vol. 48, no. 4, pp. 261– 268, 2014

    work page 2014

  9. [9]

    Next generation of computer vision for plant disease monitoring in precision agriculture: A contemporary survey, taxonomy, experi- ments, and future direction,

    W. Ding, M. Abdel-Basset, I. Alrashdi, and H. Hawash, “Next generation of computer vision for plant disease monitoring in precision agriculture: A contemporary survey, taxonomy, experi- ments, and future direction,”Inf. Sci., vol. 665, p. 120338, 2024

    work page 2024

  10. [10]

    Advances in Deep Learning Applications for Plant Disease and Pest Detection: A Review,

    S. Wanget al., “Advances in Deep Learning Applications for Plant Disease and Pest Detection: A Review,”Remote Sens., vol. 17, no. 4, p. 698, 2025

    work page 2025

  11. [11]

    Deep Learning YOLO-Based Solution for Grape Bunch Detection and Assessment of Biophysical Lesions,

    I. Pinheiroet al., “Deep Learning YOLO-Based Solution for Grape Bunch Detection and Assessment of Biophysical Lesions,” Agronomy, vol. 13, no. 4, p. 1120, 2023

    work page 2023

  12. [12]

    Deep learning models for plant disease detec- tion and diagnosis,

    K. P. Ferentinos, “Deep learning models for plant disease detec- tion and diagnosis,”Comput. Electron. Agr ., vol. 145, pp. 311– 318, 2018

    work page 2018

  13. [13]

    Recent advances in plant disease detection: Challenges and opportunities,

    M. Shafay, T. Hassan, M. Owais, I. Hussain, S. G. Khawaja, L. Seneviratne, and N. Werghi, “Recent advances in plant disease detection: Challenges and opportunities,”Plant Methods, vol. 21, no. 1, p. 140, 2025

    work page 2025

  14. [14]

    Factors influencing the use of deep learning for plant disease recognition,

    J. G. A. Barbedo, “Factors influencing the use of deep learning for plant disease recognition,”Biosyst. Eng., vol. 172, pp. 84–91, 2018

    work page 2018

  15. [15]

    R. C. Gonzalez and R. E. Woods,Digital Image Processing. Prentice Hall, 2008

    work page 2008

  16. [16]

    Rethink- ing Image Histogram Matching for Image Classification,

    R. Otsuka, Y . Shoji, Y . Ogino, T. Toizumi, and A. Ito, “Rethink- ing Image Histogram Matching for Image Classification,” 2025

    work page 2025

  17. [17]

    Deep Neural Networks Based Recognition of Plant Diseases by Leaf Image Classification,

    S. Sladojevic, M. Arsenovic, A. Anderla, D. Culibrk, and D. Stefanovic, “Deep Neural Networks Based Recognition of Plant Diseases by Leaf Image Classification,”Comput. Intell. Neurosci., vol. 2016, no. 1, p. 3289801, 2016

    work page 2016

  18. [18]

    Comparative performance of four CNN-based deep learning variants in detecting Hispa pest, two fungal diseases, and NPK deficiency symptoms of rice (Oryza sativa),

    B. Dey, M. Masum Ul Haque, R. Khatun, and R. Ahmed, “Comparative performance of four CNN-based deep learning variants in detecting Hispa pest, two fungal diseases, and NPK deficiency symptoms of rice (Oryza sativa),”Comput. Electron. Agr ., vol. 202, p. 107340, 2022

    work page 2022

  19. [19]

    Multi-granularity Feature Ex- traction Based on Vision Transformer for Tomato Leaf Disease Recognition,

    S. Wu, Y . Sun, and H. Huang, “Multi-granularity Feature Ex- traction Based on Vision Transformer for Tomato Leaf Disease Recognition,” in2021 3rd Int. Acad. Exch. Conf. Sci. Technol. Innov. IAECST, 2021, pp. 387–390

    work page 2021

  20. [20]

    Comparison of Deep Neural Networks in Detecting Field Grapevine Diseases Using Transfer Learning,

    A. Morellos, X. E. Pantazi, C. Paraskevas, and D. Moshou, “Comparison of Deep Neural Networks in Detecting Field Grapevine Diseases Using Transfer Learning,”Remote Sens., vol. 14, no. 18, p. 4648, 2022

    work page 2022

  21. [21]

    Early detection of downy mildew in vineyards using deep neural networks for semantic segmentation,

    I. Hern ´andez, R. Silva, P. Melo-Pinto, S. Guti ´errez, and J. Tardaguila, “Early detection of downy mildew in vineyards using deep neural networks for semantic segmentation,”Biosyst. Eng., vol. 252, pp. 15–31, 2025

    work page 2025

  22. [22]

    Using Deep Learning for Image-Based Plant Disease Detection,

    S. P. Mohanty, D. P. Hughes, and M. Salath ´e, “Using Deep Learning for Image-Based Plant Disease Detection,”Front. Plant Sci., vol. 7, p. 1419, 2016

    work page 2016

  23. [23]

    A review on the main challenges in automatic plant disease identification based on visible range images,

    J. G. A. Barbedo, “A review on the main challenges in automatic plant disease identification based on visible range images,” Biosyst. Eng., vol. 144, pp. 52–60, 2016

    work page 2016

  24. [24]

    Super-Resolution of Sentinel-2 Images Using Convolutional Neural Networks and Real Ground Truth Data,

    M. Galar, R. Sesma, C. Ayala, L. Albizua, and C. Aranda, “Super-Resolution of Sentinel-2 Images Using Convolutional Neural Networks and Real Ground Truth Data,”Remote Sens., vol. 12, no. 18, p. 2941, 2020

    work page 2020

  25. [25]

    Restoring Images in Adverse Weather Conditions via Histogram Trans- former,

    S. Sun, W. Ren, X. Gao, R. Wang, and X. Cao, “Restoring Images in Adverse Weather Conditions via Histogram Trans- former,” inECCV 2024. Cham: Springer, 2025, pp. 111–129

    work page 2024

  26. [26]

    A survey on Image Data Augmentation for Deep Learning,

    C. Shorten and T. M. Khoshgoftaar, “A survey on Image Data Augmentation for Deep Learning,”J. Big Data, vol. 6, no. 1, p. 60, 2019

    work page 2019

  27. [27]

    Improving Deep Learning with Generic Data Augmentation,

    L. Taylor and G. Nitschke, “Improving Deep Learning with Generic Data Augmentation,” in2018 IEEE Symp. Ser. Comput. Intell., 2018, pp. 1542–1547

    work page 2018

  28. [28]

    Data Augmentation by Pairing Samples for Images Classification,

    H. Inoue, “Data Augmentation by Pairing Samples for Images Classification,”ArXiv, 2018

    work page 2018

  29. [29]

    Random Erasing Data Augmentation,

    Z. Zhong, L. Zheng, G. Kang, S. Li, and Y . Yang, “Random Erasing Data Augmentation,”Proc. AAAI Conf. Artif. Intell., vol. 34, no. 07, pp. 13 001–13 008, 2020

    work page 2020

  30. [30]

    GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks,

    C. Bowleset al., “GAN Augmentation: Augmenting Training Data using Generative Adversarial Networks,”ArXiv, 2018

    work page 2018

  31. [31]

    LabelImg,

    Tzutalin, “LabelImg,” 2015

    work page 2015