Optimized imaging prefiltering for enhanced image segmentation

Felipe Osorio; Grisel Britos; Ronny Vallejos; Sebastian Vidal

REVIEW 5 major objections 6 minor 32 references

Optimized imaging prefiltering for enhanced image segmentation

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Box-Cox prefiltering with a likelihood-estimated parameter improves segmentation for classical discriminant models, while deep neural networks show no consistent gain.

desk verdict The paper's headline LDA/QDA improvement doesn't survive its own tables; the metric-optimal lambda observation is the real takeaway. read the letter →

arxiv 2508.03653 v1 pith:BI6G7L65 submitted 2025-08-05 stat.AP stat.ME

Ronny Vallejos , Felipe Osorio , Sebastian Vidal , Grisel Britos This is my paper

classification stat.APstat.ME MSC 62H3068U10

keywords Box-Coxtransformationimagesegmentationprefilteringdiscriminantanalysismachinelearningdeepfeatureseparabilitycomputationalefficiency

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

The paper argues that applying a Box-Cox power transformation as a prefilter, with the transformation parameter $\lambda$ estimated from a profile likelihood, improves image segmentation for traditional statistical classifiers. This matters because segmentation problems with little or no labeled data, such as crack detection or lunar rock identification, give an edge to models like linear and quadratic discriminant analysis that assume normality and stable variance. On a concrete crack image, LDA and QDA improved by about three percentage points in precision, recall, and F1 after prefiltering; on lunar rock images, QDA's correct classification of rock pixels went from zero to 79 percent. The same experiments on U-Net, FCN, and DeepLab trained on satellite images showed no consistent improvement, and deep models often lost Dice and IoU points. If the claim is right, cheap assumption-based classifiers become more attractive for low-label segmentation, with training times measured in seconds rather than hours.

What carries the argument

The central object is the Box-Cox-transformed, vectorized pixel intensity $y(\lambda)$, with $y_i^{(\lambda)} = (y_i^\lambda-1)/\lambda$ for $\lambda \neq 0$ and $\log y_i$ for $\lambda = 0$, followed by histogram stretching to the full gray-level range. The parameter $\lambda$ is estimated by maximizing the profile likelihood of the assumed model $y(\lambda) \sim N(A\theta,\sigma^2 I)$; that estimated $\lambda$ is what makes the prefilter "optimized." The mechanism is distributional: compressing skew and stabilizing variance makes the class-conditional Gaussian assumptions of LDA and QDA fit the pixel intensities better, which sharpens the linear and quadratic decision boundaries.

What would settle it

Take a set of labeled images with ground-truth masks and compare LDA/QDA segmentation under histogram stretching only, Box-Cox with profile-likelihood $\lambda$, and Box-Cox with $\lambda$ tuned to maximize $\kappa$ on a validation split; if likelihood-based $\lambda$ does not beat stretching alone, or if validation-tuned $\lambda$ beats it, the paper's recommendation fails. Also check whether the unshifted Box-Cox transformation is computable on images containing zero-intensity pixels.

Watch

Extended reading notes

Core claim

The central claim is that the Box-Cox transformation, applied to grayscale pixel intensities before segmentation and then followed by histogram stretching, enhances feature separability for models whose decision rules assume normal classes and stable variance. The paper estimates $\lambda$ by maximizing the profile likelihood of the model $y(\lambda) \sim N(A\theta,\sigma^2 I)$ on the vectorized pixels. Experiments show that likelihood-selected $\lambda$ improves LDA and QDA segmentation of a concrete crack image and a lunar rock image, including a correction of rock-pixel classification from entirely wrong to 79 percent correct, while the same prefiltering does not consistently help U-Net, FCN, or DeepLab on satellite water images. The paper also notes that the likelihood-optimal $\lambda$ is not always the $\lambda$ that maximizes precision or the kappa coefficient, so parameter choice involves a trade-off between concordance and accuracy.

Load-bearing premise

The load-bearing premise is that the profile-likelihood model used to choose $\lambda$, a linear Gaussian model that treats each pixel as an independent observation with the same variance, is a faithful enough description of images, and that the transformation can be applied to the zero-valued intensities that real grayscale images contain; if spatial correlation or zero pixels change the estimate of $\lambda$ materially, the "optimal" prefiltering is not well founded.

Editorial extensions

If this is right

LDA and QDA can reach near-SVM segmentation quality on crack images after Box-Cox prefiltering, while cutting training time from over a hundred seconds to under a second.
On lunar rock images, the prefilter lifted QDA's rock-class recall from a completely wrong classification to 79 percent correct, making a three-class segmentation problem tractable without a neural network.
Deep learning pipelines should not expect accuracy gains from Box-Cox prefiltering; in the reported satellite experiments U-Net, FCN, and DeepLab stayed flat or declined.
The likelihood-based $\lambda$ is a useful default, but practitioners should treat it as a starting point and inspect precision and $\kappa$ across a grid of $\lambda$ values.
The approach offers a practical path for segmentation in resource-constrained or low-label settings, where classical distributional classifiers become competitive because they do not require large training sets.

Reading between the lines

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

The paper does not test $\lambda$ selection against a labeled validation set; given the reported mismatch between likelihood-optimal and concordance-optimal $\lambda$, tuning $\lambda$ to maximize Dice or $\kappa$ on a small validation split is a direct testable extension.
In very-low-label regimes, the cheap prefilter could be combined with self-training or pseudo-labeling, where confident LDA/QDA predictions on unlabeled images refine the training set; this is not explored.
The deep-model result may depend on the amount of training data; with severely reduced training sets, prefiltering could help neural models too, since their internal contrast normalization would have less data to learn from.

Share X Bluesky LinkedIn Reddit HN

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

Short version: the paper's central claim about LDA/QDA improvements is not supported by its own tables. Table 2 reports LDA F1 = 97 before transformation, but Table 3 shows the same pre-transformation LDA assigns 100% of true crack pixels to concrete — recall 0, F1 0. The text even says the original segmentation was entirely black. You can't get 96/98/97 from that confusion matrix. So the "approximately 3% improvement" is an artifact, and the paper's main evidence collapses.

That said, the paper isn't worthless. The observation that likelihood-optimal λ differs from metric-optimal λ (Section 3.3, Figures 5–7) is genuinely interesting and worth following up. The authors also ship code and data, and they are candid about spatial autocorrelation being future work. The negative result on deep models is plausible, though undercut by the fact that the networks were trained on original images and only tested on transformed ones — that's not a fair comparison.

Other soft spots, in order of severity. The F1 formula in Section 2.2 is missing the factor 2; as written, Precision·Recall/(Precision+Recall) is not F1. The Box-Cox setup assumes positive y, but grayscale pixels include zeros; the paper never states the shift c used in practice. And the lunar rock result is a mixed bag: QDA rock recall jumps from 0 to 79, but sky accuracy drops, so the "beneficial" story is selective. The computational efficiency point (Table 4) is fine but trivial: LDA/QDA are closed-form, of course they're faster than SVM.

Bottom line: there's a publishable kernel here — a careful study of Box-Cox prefiltering for segmentation, with the metric-optimal λ finding — but this version has a load-bearing inconsistency. I wouldn't desk-reject it outright; I'd send it to a referee with instructions to check the confusion-matrix math. But in current form I would not cite it for the quantitative claims.

Referee Report

5 major / 6 minor

Summary. The paper proposes using the Box-Cox transformation with a likelihood-estimated parameter λ as a preprocessing step for image segmentation, applied to grayscale images followed by histogram stretching. The authors compare classical machine learning methods (SVM, LightGBM, LDA, QDA, KNN) and deep learning models (DeepLabV3, U-Net, FCN) on a crack image, a lunar rock image, and satellite water images. The central claim is that the transformation improves segmentation for discriminant analysis in low-label settings while not consistently helping deep models. The paper also discusses the trade-off between likelihood-based and metric-based choices of λ.

Significance. If the central claim were established, the paper would offer a simple, computationally cheap preprocessing step with a statistically motivated parameter choice that benefits classical segmentation methods in data-scarce settings. The authors make their code, data, and trained weights publicly available, which is a strength. However, the current evidence is not reliable: the principal quantitative demonstration (the crack image) is internally inconsistent, the statistical derivation of λ contains sign errors, and the deep-learning comparison is not a fair test. The idea is plausible, but the manuscript as written does not support its headline conclusions. With substantial corrections and additional experiments, the approach could be a useful contribution, but the present version requires major revision.

major comments (5)

[Section 3.2, Tables 2 and 3] Table 3 shows that before Box-Cox, LDA assigns 0% of true crack pixels to the crack class (Concrete 100, Crack 0), consistent with the text's statement that the original segmentation yielded entirely black images. Under standard definitions, recall is then 0 and F1 is 0, yet Table 2 reports LDA recall of 98 and F1 of 97 before the transformation. This internal contradiction invalidates the claimed 'approximately 3%' improvement and calls into question every baseline metric in Table 2. Because this single crack image is the main evidence for the benefit of the prefilter on LDA/QDA, the central quantitative result is unsupported as reported.
[Section 2.1, Eq. (1)] The Box-Cox transformation in Eq. (1) is defined only for positive observations, but the text states that y contains nonnegative pixel intensities. Grayscale images contain zero-valued pixels, and the shift parameter c from Eq. (2) is never specified for the experiments. Consequently, the likelihood and the resulting λ estimates are undefined for zero pixels, which affects all λ values used in Sections 3.1–3.3. The paper must either justify the use of a shift or describe how zeros are handled.
[Section 3.1] The deep-learning comparison is flawed. The networks were trained exclusively on original images and then evaluated on Box-Cox-transformed test images, as stated in Section 3.1: 'After training all models and tuning the hyperparameters, the test set images were segmented both before and after the Box-Cox transformation.' Any performance difference could be due to the distribution shift between training and test inputs rather than to the transformation itself. A valid comparison would require retraining or fine-tuning the models on the transformed images.
[Section 2.1, profile likelihood] The Gaussian model for the transformed data is underspecified: the design matrix A is not defined, and the assumption of independent errors (σ²I) ignores the strong spatial autocorrelation in images, which the authors acknowledge in Section 4 as future work. Without a precise specification of A and a treatment of dependence, the profile-likelihood estimates of λ are not reproducible and their statistical justification is incomplete, even setting aside the sign errors noted below.
[Section 2.1, likelihood equations] The displayed log-likelihood and profile likelihood contain sign errors. The standard log-likelihood for the original data is −(n/2)log(σ²) − RSS/(2σ²) + (λ−1)Σ log(y_i), but the manuscript has +(n/2)log(σ²) − RSS/(2σ²) − (λ−1)Σ log(y_i). The subsequent profile likelihood changes the sign of the Jacobian term, giving ℓ(λ) = (n/2)log(RSS/n) + (λ−1)Σ log(y_i) instead of the correct −(n/2)log(RSS/n) + (λ−1)Σ log(y_i). Maximizing the displayed profile likelihood would tend to maximize the residual sum of squares, the opposite of the intended estimation. This error undermines the λ estimates used in all experiments.

minor comments (6)

[Section 2.2] The F1 score formula is missing the factor 2: it should be F1 = 2·Precision·Recall / (Precision + Recall).
[Section 2.1, after bθ] The estimate of σ² is written as (1/n)||y(λ) − A bθ||; it should be the squared norm, (1/n)||y(λ) − A bθ||².
[Section 3.3, Figure 6] The text states that the λ maximizing concordance and precision is 0.31, but the caption of Figure 6 reports 0.26; these should be reconciled.
[Section 3.3] The sentence 'the maximum and minimum of each curve correspond to the minimum and maximum of the other' is self-contradictory; clarify whether the extrema coincide or are inverted.
[Table 1 and throughout] The tables use commas as decimal separators (e.g., 44,77) inconsistently; use decimal points for clarity.
[References] Reference [29] contains a leading anomalous dot and formatting issue ('． Vallejos, R.'); also, 'Tables 3' in Section 3.2 should be 'Table 3'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: lambda is estimated by likelihood and segmentation metrics are evaluated afterward, with Section 3.3 explicitly showing the likelihood-optimal lambda is not metric-optimal.

full rationale

The paper's central derivation is not circular. The transformation parameter lambda is estimated from a profile likelihood under a linear Gaussian model for transformed pixel intensities (Section 2.1), entirely separate from the segmentation quality metrics. The results in Section 3 are computed by applying the estimated lambda as a fixed prefilter and then evaluating classifiers; no segmentation metric is used to fit lambda. Section 3.3 even demonstrates that the likelihood-optimal lambda (0.43 for the crack image) is neither the precision-optimal nor the kappa-optimal value, which is the opposite of a fitted-input-renamed-as-prediction pattern. The self-citations in the paper ([2], [20], [29]) appear only in the discussion of future concordance measures and are not load-bearing inputs to any result. The Table 2/3 discrepancy (LDA recall 98 in Table 2 versus a Crack=0 row in Table 3) is an internal consistency/correctness concern, not a circularity of derivation; likewise the unspecified shift c for zero pixel intensities is a modeling gap, not a definitional circularity. Hence no circular step is exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method relies on a likelihood model for pixel intensities, an unstated design matrix, and single-image experiments; no new theoretical entities are introduced.

free parameters (3)

Box-Cox parameter lambda = 0.43 (crack), 2.54 (lunar rock)
Estimated by profile likelihood from each image; the paper's method depends on it, and Section 3.3 shows the estimate is not optimal for the segmentation metrics.
Design matrix A in Gaussian model = unspecified
The likelihood uses y(lambda) ~ N(A theta, sigma^2 I) but the paper never states what A contains (intercept, pixel coordinates, etc.), so lambda estimates are conditional on an unstated modeling choice.
Decision threshold 0.5 = 0.5
Applied to convert probabilistic neural network outputs into water/non-water labels in Section 3.1.

assumptions (4)

domain assumption Vectorized pixel intensities after Box-Cox follow a linear Gaussian model with iid errors
Section 2.1 states y(lambda) ~ N(A theta, sigma^2 I); pixel intensities are spatially correlated, so the iid assumption is questionable and the paper itself lists spatial autocorrelation as future work.
domain assumption A is a full-rank design matrix of dimension n x p
Section 2.1 assumes A without specifying its construction, making the profile likelihood and lambda dependent on an unstated modeling choice.
domain assumption Box-Cox transformation is applicable to all grayscale intensities, including zeros
The paper defines y in R_+^n and uses Eq. (1), which requires positive values; handling of zero-valued pixels is not described.
domain assumption Ground-truth masks are correct
All metrics compare predictions against masks from the Kaggle and Keio datasets; no quality assessment of masks is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Optimized imaging prefiltering for enhanced image segmentation." pith.science (2026). https://pith.science/paper/BI6G7L65

@misc{pith2026250803653,
  author       = {Pith},
  title        = {Pith review of: Optimized imaging prefiltering for enhanced image segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BI6G7L65}},
  note         = {Machine review of arXiv:2508.03653}
}

read the original abstract

The Box-Cox transformation, introduced in 1964, is a widely used statistical tool for stabilizing variance and improving normality in data analysis. Its application in image processing, particularly for image enhancement, has gained increasing attention in recent years. This paper investigates the use of the Box-Cox transformation as a preprocessing step for image segmentation, with a focus on the estimation of the transformation parameter. We evaluate the effectiveness of the transformation by comparing various segmentation methods, highlighting its advantages for traditional machine learning techniques-especially in situations where no training data is available. The results demonstrate that the transformation enhances feature separability and computational efficiency, making it particularly beneficial for models like discriminant analysis. In contrast, deep learning models did not show consistent improvements, underscoring the differing impacts of the transformation across model types and image characteristics.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 31 canonical work pages

[1]

J., Doksum, K

Bickel, P. J., Doksum, K. A. (1981). An analysis of transformations revisited. Journal of the American Statistical Association 7, 296–311

work page 1981
[2]

Acosta, J., Vallejos, R., Osorio, F., Ellison, A., de Castro, M. (2024). Comparing two spatial variables with the probability of agreement. Biometrics 80, ujae009

work page 2024
[3]

E., Cox, D

Box, G. E., Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society: Series B 26, 211–243

work page 1964
[4]

Cheddad, A. (2020). On box-cox transformation for image normality and pattern classification. IEEE Access 8, 154975–154983

work page 2020
[5]

A., Stephens, M

Chen, G., Lockhart, R. A., Stephens, M. A. (2002) Box–Cox transformations in linear models: large sample theory and tests of normality (with discussion). Canadian Journal of Statistics 30,177—234

work page 2002
[6]

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018). Encoder- Decoder with Atrous Separable Convolution for Semantic Image Segmentation. arXiv:1802.02611 [cs.CV]. https://doi.org/10.48550/arXiv.1802.02611

arXiv doi:10.48550/arxiv.1802.02611 2018
[7]

Christudhas, C., Fathima, A. (2025). Results in Engineering 25, 104481

work page 2025
[8]

Cressie, N. (1993). Statistics for Spatial Data (Revised Edition). John Wiley & Sons, Inc

work page 1993

Show all 32 references

[9]

de la Cruz Huayanay, A., Baz´ an, J., Russo, C. M. (2025). Performance of evalu- ation metrics for classification in imbalanced data. Computational Statistics, 40 1447–1473

2025
[10]

X., Huang, D

Du, J. X., Huang, D. S., Wang, X. F., Gu, X. (2007). Shape recognition based on neural networks trained by differential evolution algorithm, Neurocomputing 70, 896–903

2007
[11]

Garcia-Lamont, F., Cervantes, J., L´ opez, A., Rodriguez. L. (2018). Segmentation of images by color features: A survey. Neurocomputing 192, 1–27

2018
[12]

Woods, and S.L

Gonzalez, R.C., R.E. Woods, and S.L. Eddins, Digital image processing using MATLAB. 2011: Tata McGraw-Hill Education Private Ltd

2011
[13]

Jakhar, K., Kaur, A., and Gupta, M. (2021). Pneumothorax segmentation: Deep learning image segmentation to predict pneumothorax. arXiv:1912.07329

2021 arXiv
[14]

A., Draper, N

John, J. A., Draper, N. R. (1980). An alternative of family transformations, Applied Statistics 29, 190–197

1980
[15]

Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., Liu, T.-Y. (2017). LightGBM: A Highly Efficient Gradient Boosting Decision Tree. Advances in Neural Information Processing Systems, 30, 3146–3154

2017
[16]

Landis, J. R. Koch, G. G. (1977). The measurement of observer agreement for categorical data. Biometrics 33, 159–174

1977
[17]

Manly, B. F. (1976). Exponential data transformations. The Statistician 25, 37– 42

1976
[18]

K., Abonyi, J

Maurya, L., Lohchab, V., Mahapatra, P. K., Abonyi, J. (2022). Contrast and brightness balance in image enhancement using Cuckoo Search-optimized image 19 fusion. Journal of King Saud University - Computer and Information Sciences 34, 7247–7258

2022
[19]

McCullagh, P. (2002). Comment on Box–Cox transformations in linear models: large sample theory and tests of normality by Chen, Lockhart and Stephens. Canadian Journal of Statistics 30, 212–213

2002
[20]

, Bustos, O

Ojeda, S., Vallejos, R. , Bustos, O. (2010). A new image segmentation algorithm with applications to image inpainting. Computational Statistics & Data Analysis 54, 2082-2093

2010
[21]

Improving your data transformations: Applying the Box-Cox transformation, Practical Assessment, Research, and Evaluation 15(1): 12

Osborne, J., (2010). Improving your data transformations: Applying the Box-Cox transformation, Practical Assessment, Research, and Evaluation 15(1): 12. doi: https://doi.org/10.7275/qbpc-gk17

2010 doi
[22]

and Sorguc, A

Ozgenel, C. and Sorguc, A. G. (2018). Performance comparison of pretrained convolutional neural networks on crack detection in buildings. In Proceedings of the 35th International Symposium on Automation and Robotics in Construction (ISARC)

2018
[23]

R., Pal, S

Pal, N. R., Pal, S. K. (1993). A review on image segmentation techniques. Pattern Recognition 26, 1274– 1294

1993
[24]

Petrakis, G., and Partslnevelos, P. (2024). Lunar ground segmentation using a modified U-net neural network. Machine Vision and Applications 35, 50

2024
[25]

Proietti T., Riani, M. (2009). Seasonal adjustment and transformations. Journal of Time Series Analysis 30,47–69

2009
[26]

C., Corbellini, A

Riani, M., Atkinson, A. C., Corbellini, A. (2023). Automatic robust Box–Cox and extended Yeo–Johnson transformations in regression. Statistical Methods and Applications 32, 75–102

2023
[27]

Ronneberger, O., Fischer, P., Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer- Assisted Intervention (MICCAI), 9351, 234—241

2015
[28]

Sim, J., Wright, C. (2005). The kappa statistic in reliability studies: Use, interpre- tation, and sample size requirements. Journal of the American Physical Therapy 85, 257–268

2005
[29]

, P´ erez, J., Ellison, A., Richardson, A

Vallejos, R. , P´ erez, J., Ellison, A., Richardson, A. (2020). A spatial concordance correlation coefficient with an application to image analysis. Spatial Statistics 40, 100405

2020
[30]

Yeo, I., Johnson, R. (2000). A new family of power transformations to improve normality or symmetry. Biometrika 87, 954–959

2000
[31]

Yu, Y., Wang, C., Fu, Q., Kou, R., Huang, F., Yang, B., Yang, T., Gao, M. (2023). Techniques and Challenges of Image Segmentation: A Review. Electronics 12, 1199

2023
[32]

Zhou, He., Zou, H. (2024). A Non-Parametric Box-Cox Approach to Robustifying High-Dimensional Linear Hypothesis Testing. arXiv:2405.12816. https://doi.org/ 10.48550/arXiv.2405.12816. 20

2024 doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

J., Doksum, K

Bickel, P. J., Doksum, K. A. (1981). An analysis of transformations revisited. Journal of the American Statistical Association 7, 296–311

work page 1981

[2] [2]

Acosta, J., Vallejos, R., Osorio, F., Ellison, A., de Castro, M. (2024). Comparing two spatial variables with the probability of agreement. Biometrics 80, ujae009

work page 2024

[3] [3]

E., Cox, D

Box, G. E., Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical Society: Series B 26, 211–243

work page 1964

[4] [4]

Cheddad, A. (2020). On box-cox transformation for image normality and pattern classification. IEEE Access 8, 154975–154983

work page 2020

[5] [5]

A., Stephens, M

Chen, G., Lockhart, R. A., Stephens, M. A. (2002) Box–Cox transformations in linear models: large sample theory and tests of normality (with discussion). Canadian Journal of Statistics 30,177—234

work page 2002

[6] [6]

Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., & Adam, H. (2018). Encoder- Decoder with Atrous Separable Convolution for Semantic Image Segmentation. arXiv:1802.02611 [cs.CV]. https://doi.org/10.48550/arXiv.1802.02611

arXiv doi:10.48550/arxiv.1802.02611 2018

[7] [7]

Christudhas, C., Fathima, A. (2025). Results in Engineering 25, 104481

work page 2025

[8] [8]

Cressie, N. (1993). Statistics for Spatial Data (Revised Edition). John Wiley & Sons, Inc

work page 1993