Performance Evaluation of Image Enhancement Techniques on Transfer Learning for Touchless Fingerprint Recognition

Alavikunhu Panthakkan; Dilavar P D; Saad Ali Amin; S M Anzar; S Sreehari

REVIEW 4 major objections 5 minor 12 references

Performance Evaluation of Image Enhancement Techniques on Transfer Learning for Touchless Fingerprint Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Applying a fixed image-enhancement pipeline before transfer learning raises touchless fingerprint recognition accuracy, with VGG-16 reaching 98% (93% without), and the same trend across VGG-19, ResNet-50, and Inception-V3.

desk verdict A standard benchmark whose central comparison is confounded by different dataset sizes and an unspecified split; the preprocessing gain is plausible but not established. read the letter →

arxiv 2502.04680 v1 pith:PXRDIJVC submitted 2025-02-07 cs.CV cs.LG

S Sreehari , Dilavar P D , S M Anzar , Alavikunhu Panthakkan , Saad Ali Amin This is my paper · ORCID

classification cs.CVcs.LG

keywords touchlessfingerprintrecognitiontransferlearningimageenhancementVGG-16CLAHESIFTbiometricauthenticationIIT-Bombaydatabase

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 tries to establish that a fixed image-enhancement pipeline applied to touchless fingerprint photos makes transfer-learned CNN classifiers markedly more accurate. Using the IIT-Bombay Touchless and Touch-Based Fingerprint Database (200 subjects, 200 classes), the authors compare four pretrained architectures with and without preprocessing. In their experiments, every model improves after enhancement: VGG-16 goes from 93% to 98% accuracy, VGG-19 from 92% to 97%, ResNet-50 from 64% to 76%, and Inception-V3 from 17% to 64%. If that is right, simple preprocessing is a high-value addition to contactless biometric systems, which otherwise suffer from low-contrast, noisy images.

What carries the argument

The load-bearing mechanism is a fixed chain of operations applied to each touchless fingerprint before classification: normalization; contrast-limited adaptive histogram equalization (CLAHE); SIFT keypoint detection; thresholding; cropping; a Laplacian filter; inversion; sharpening; contrast adjustment; and dilation. These steps are intended to turn low-contrast, noisy touchless photos into images whose ridge and valley structure is clear enough for a pretrained CNN to classify. The classifiers are VGG-16, VGG-19, ResNet-50, and Inception-V3, fine-tuned on the IIT-Bombay touchless set with categorical crossentropy and the Nadam optimizer.

What would settle it

Run both pipelines on a strict subject-disjoint split with equal numbers of training images per condition; if VGG-16's 98%-versus-93% gap collapses, the claimed benefit of preprocessing is not supported.

Watch

Extended reading notes

Core claim

The central claim is that 'indirect' recognition—enhancing images before feeding them to a pretrained network—outperforms 'direct' recognition on raw images across all four tested architectures. The flagship result is VGG-16: 98% accuracy with the preprocessing chain versus 93% without, and the paper reports the same ordering for VGG-19 (97% versus 92%), ResNet-50 (76% versus 64%), and Inception-V3 (64% versus 17%). The authors interpret this as evidence that preprocessing clarifies ridge structure enough for transfer-learned features to be useful, making touchless fingerprint identification practical with standard CNN backbones.

Load-bearing premise

The comparison assumes that the training and test partitions are valid and comparable, meaning no person's fingerprint images appear in both, and that the different training-set sizes (3200 versus 1800) do not explain the accuracy gain.

Editorial extensions

If this is right

A touchless fingerprint system can reach 98% accuracy with a standard pretrained network plus the preprocessing chain, without needing a custom architecture.
Every architecture tested improves after preprocessing, so the benefit is not specific to VGG-16; the same pipeline can lift weaker backbones such as ResNet-50 and Inception-V3.
On raw images, Inception-V3 nearly fails at 17% accuracy, suggesting that some models may depend on preprocessing to learn anything useful.
Combining preprocessing with transfer learning makes contactless biometric identification feasible with modest hardware and small datasets.

Reading between the lines

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

Beyond the paper's claims, the two conditions are not matched: 3200/1600 images without preprocessing versus 1800/600 with it, so a matched-size comparison would isolate how much of the gain is enhancement rather than data quantity.
Beyond the paper's claims, the abstract calls VGG-16's result '98% training, 93% testing,' but the table lists 0.93 as recall; re-reading the result as 98% accuracy with 93% recall would change how the headline number is quoted.
Beyond the paper's claims, a subject-disjoint evaluation would reveal whether the models learn identity or image-level artifacts, since augmented copies of the same finger may otherwise appear in both training and testing.
Beyond the paper's claims, the fixed preprocessing chain is not ablated, so no single step (CLAHE, SIFT, Laplacian, dilation) can be credited; testing subsets would identify the active ingredient.

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

Short version: this is a straightforward benchmark, not a new method. It applies standard preprocessing (CLAHE, SIFT, thresholding, Laplacian, dilation) and standard pretrained CNNs to the IIT Bombay touchless dataset, and reports that preprocessing helps. The trend is consistent across VGG-16, VGG-19, ResNet-50, and Inception-V3, and the paper gives full metric tables and accuracy/loss curves. That is the useful part. Credit where earned: it uses a public dataset, describes the pipeline step by step, and reports per-model precision, recall, and F1, which is more than many papers at this level do. The self-citations are background and not load-bearing, so I don't count them against the work. The problem is that the experiment cannot support the causal claim. The two arms use different training/test sizes: 3,200/1,600 without preprocessing versus 1,800/600 with preprocessing. Training-set size and augmentation ratio are therefore conflated with preprocessing. The paper also never states whether the train/test split is subject-disjoint. With 200 subjects and four samples each, an image-level split can put the same identity in both training and testing. The jumps for ResNet-50 (0.64 to 0.76) and Inception-V3 (0.17 to 0.64) are large enough that leakage or dataset-size effects are a real alternative explanation. No error bars, no repeated runs, no code, no preprocessing parameter values. The abstract also says VGG-16 achieved 98% in training and 93% in testing, which conflicts with Table II where 93% is the no-preprocessing accuracy; that is a reporting error a referee would catch. I don't think the authors are hiding anything; the paper is just under-specified. The central claim that preprocessing helps touchless fingerprint CNNs is plausible and consistent with prior work they cite. But this manuscript, as written, does not add solid evidence beyond a single run on one public dataset. Who is this for? Someone wanting a quick sanity check that standard preprocessing plus transfer learning works on touchless fingerprints might skim it. It is not a paper I would cite or build on. In its current form I would desk reject it; if the authors redo the experiment with matched dataset sizes, a subject-disjoint split, repeated runs, and shipped code/parameters, the comparison could become a useful workshop-level or short-report contribution.

Referee Report

4 major / 5 minor

Summary. The manuscript empirically compares two pipelines for touchless fingerprint recognition on the IIT-Bombay Touchless and Touch-Based Fingerprint Database: direct classification with transfer learning (no preprocessing) versus classification after a multi-stage enhancement pipeline (normalization, CLAHE, SIFT, thresholding, cropping, Laplacian filtering, inversion, sharpening, contrast, dilation). Four pre-trained CNNs (VGG-16, VGG-19, ResNet-50, Inception-V3) are evaluated in terms of accuracy, loss, precision, recall, and F1-score. The paper reports consistent gains for all models with preprocessing, with VGG-16 improving from 93% to 98% accuracy, and concludes that image enhancement significantly improves transfer learning for touchless fingerprints. The manuscript does not provide code or reproducibility artifacts.

Significance. If the claimed causal effect were established, the result would be practically useful: it would show that a relatively simple, deterministic enhancement pipeline can improve transfer-learning accuracy for touchless fingerprint images across several standard CNN backbones. The paper uses a public benchmark dataset and reports per-model metrics, which are both strengths. However, the current experimental design does not yet support the headline claim, because the two compared arms differ in more than the preprocessing variable and the possibility of train/test leakage is not addressed. The manuscript is therefore better viewed as a promising preliminary study than as a definitive evaluation.

major comments (4)

[Section V.A] The comparison arm sizes are confounded with the preprocessing variable. The text states that without preprocessing the dataset was augmented to 3,200 training and 1,600 test images, whereas with preprocessing it contained 1,800 training and 600 test images. Thus the two arms differ in training-set size, test-set size, and augmentation ratio, in addition to the preprocessing pipeline. The accuracy differences in Tables I and II (e.g., ResNet-50 from 0.64 to 0.76 and Inception-V3 from 0.17 to 0.64) cannot be attributed solely to enhancement. A valid comparison requires matching the number of training and test samples across arms (or otherwise controlling for dataset size), and the augmentation budgets should be identical.
[Section V.A] The paper never states whether the train/test split is subject-disjoint. The base dataset contains 200 subjects with four touchless samples per subject; if the split is performed at the image level, the same subject can occur in both training and testing, and augmented or preprocessed versions of the same original image may cross the boundary. Such leakage can inflate the reported accuracies by letting the model memorize identities rather than learn generalizable features. The split rule must be stated explicitly, and if leakage is possible, the experiments should be repeated with a subject-disjoint split.
[Abstract, Section V.C] There is an internal contradiction in the headline results. The abstract states that VGG-16 achieved 'an accuracy of 98% in training and 93% in testing when using the enhanced images,' but Section V.C and Table II report 98% accuracy for the preprocessed VGG-16 and 93% for the no-preprocessing VGG-16. This is not a mere wording issue, because the conclusion repeats that VGG-16 achieved 98% with preprocessing and 93% without preprocessing. The abstract appears to misattribute the 93% figure to the preprocessed test set, which would contradict the central claim. The reported numbers must be made internally consistent.
[Section V.B] The claim that preprocessing 'significantly' improves performance is not supported by statistical evidence. All metrics appear to come from a single run with no error bars, no repeated trials, and no random-seed or cross-validation information. For models such as ResNet-50 and Inception-V3, whose training is prone to variance, the observed gaps could fall within run-to-run variability. The authors should report means and standard deviations over multiple runs (or an equivalent statistical treatment) before claiming significance.

minor comments (5)

[Section IV] The preprocessing pipeline is not sufficiently parameterized for reproducibility: the CLAHE clip limit and tile grid size, binarization threshold, dilation kernel size, SIFT parameters, and augmentation settings are not specified. Please provide these values or a reference to an exact implementation.
[Keywords] The author-supplied keywords ('Cybersecurity, Machine Learning, Online Security, Phishing Detection, Threat Detection') are unrelated to the paper's content on fingerprint recognition and appear to be a template leftover. They should be replaced with relevant terms such as touchless fingerprint recognition, image enhancement, transfer learning, and CNN.
[Tables I and II] The row numbering in Table I is inconsistent: the rows are labeled 1, 2, 4, 3, and the Inception-V3 row has no number. Please renumber the rows sequentially in both tables.
[Figures 3 and 4] The figure references in Section IV ('Figure 3.6', 'Figure 3.7') appear to be section-based labels rather than sequential figure numbers. Please renumber the figures and update the in-text references.
[References] Several references have incomplete or inconsistent formatting (e.g., [4] and [5] list author names in nonstandard order, and some entries lack page ranges). Please normalize all references to the journal's style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the accuracy comparison is empirical and does not reduce to its inputs; self-citations are background only.

full rationale

This is an experimental benchmark paper rather than a mathematical derivation. The claimed result (preprocessing improves transfer-learning accuracy on touchless fingerprints) is supported by accuracy/loss tables produced by training fixed pre-trained CNN architectures on the IIT-Bombay dataset with and without an enhancement pipeline. No equation defines the outcome in terms of the input, no fitted parameter is relabeled as a prediction, and no uniqueness theorem or ansatz is imported from prior work. The only self-citations are references [1]-[3] (Anzar and Sathidevi), used in the introduction for general statements about fingerprint reliability and multibiometric fusion; they are not used to compute any of the reported accuracies, so they are not load-bearing. The main validity concerns are experimental, not circular: the 'without preprocessing' arm itself applies normalization, contrast, color, brightness, and sharpening augmentations (Section III); the two arms use different training/test counts (3,200/1,600 vs 1,800/600) and the paper does not state a subject-disjoint split (Section V.A), so leakage or dataset-size effects could confound the reported gain. These are correctness risks, not instances of a claim being equivalent, by construction or by self-citation, to its inputs. Under the circularity definition used here, the paper contains no circular step.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on assumptions about the dataset split, the transferability of ImageNet features, and the identity-preserving nature of the preprocessing steps. Many preprocessing parameters are chosen by hand and not reported, so they are free parameters with respect to reproducibility.

free parameters (6)

CLAHE clip limit and tile grid size
Chosen by hand in Section IV; not reported, directly controls contrast enhancement.
Binarization threshold
Thresholding step in preprocessing (Section IV, Figure 4(iv)) has no reported value.
Dilation kernel size
Dilation step in preprocessing (Figure 4(x)) is unspecified; affects ridge thickness.
Augmentation parameters
Section III and V.A state augmentation is used but the specific operations and magnitudes are not given.
SIFT parameters
SIFT keypoint detection is used for cropping (Section IV), but contrast threshold, edge threshold, and descriptor settings are omitted.
Training hyperparameters = epochs=100, batch_size=32, optimizer=Nadam
Chosen by hand in Section V.B; no validation-based tuning or sensitivity analysis is reported.

assumptions (4)

domain assumption ImageNet-pretrained weights transfer useful features to fingerprint images
The transfer learning setup presumes ImageNet features help fingerprint classification; this is not tested against training from scratch.
domain assumption The enhancement pipeline preserves identity-relevant ridge structure while removing background
The preprocessing steps (thresholding, cropping, dilation) could remove or distort features; the paper assumes they improve class separability.
domain assumption Train and test sets are class-disjoint
Section V.A reports 200 classes but does not state that subjects in the test set are unseen during training; no subject-level split is described.
domain assumption Augmentation does not cause leakage
Augmented images are generated from the same source images; without explicit split ordering, duplicated content may appear in both training and testing.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Performance Evaluation of Image Enhancement Techniques on Transfer Learning for Touchless Fingerprint Recognition." pith.science (2026). https://pith.science/paper/PXRDIJVC

@misc{pith2026250204680,
  author       = {Pith},
  title        = {Pith review of: Performance Evaluation of Image Enhancement Techniques on Transfer Learning for Touchless Fingerprint Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PXRDIJVC}},
  note         = {Machine review of arXiv:2502.04680}
}

read the original abstract

Fingerprint recognition remains one of the most reliable biometric technologies due to its high accuracy and uniqueness. Traditional systems rely on contact-based scanners, which are prone to issues such as image degradation from surface contamination and inconsistent user interaction. To address these limitations, contactless fingerprint recognition has emerged as a promising alternative, providing non-intrusive and hygienic authentication. This study evaluates the impact of image enhancement tech-niques on the performance of pre-trained deep learning models using transfer learning for touchless fingerprint recognition. The IIT-Bombay Touchless and Touch-Based Fingerprint Database, containing data from 200 subjects, was employed to test the per-formance of deep learning architectures such as VGG-16, VGG-19, Inception-V3, and ResNet-50. Experimental results reveal that transfer learning methods with fingerprint image enhance-ment (indirect method) significantly outperform those without enhancement (direct method). Specifically, VGG-16 achieved an accuracy of 98% in training and 93% in testing when using the enhanced images, demonstrating superior performance compared to the direct method. This paper provides a detailed comparison of the effectiveness of image enhancement in improving the accuracy of transfer learning models for touchless fingerprint recognition, offering key insights for developing more efficient biometric systems.

Figures

Figures reproduced from arXiv: 2502.04680 by the authors.

**Figure 1.** Block diagram of Touchless fingerprint Identification without Preprocessing III. BACKGROUND WORK In this approach, we expand the dataset for touchless fingerprint recognition using augmentation techniques to increase diversity and volume. Instead of traditional preprocessing, these augmentations prepare the data for training deep learning models. Geometrical augmentation is avoided to prevent distortion of fingerp… view at source ↗

**Figure 2.** (i) Input image, (ii) Normalized image, (iii) Color Enhanced image, (iv) Contrast increased image, (v) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Block diagram of Touchless fingerprint Identification with Preprocessing [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: (i) Normalization, (ii) CLAHE Enhancement Method, [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Accuracy and Loss Plots Vs Epoch of Transfer Learning Models without [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Accuracy and Loss Plots Vs Epoch of Transfer Learning Models with Preprocessing [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Optimal score level fusion using modalities reliability and separability measures,

S. Anzar and P. Sathide vi, “Optimal score level fusion using modalities reliability and separability measures,” Int. J. Comput. Appl, vol. 51, no. 16, pp. 1–8, 2012

work page 2012
[2]

On combining multi- normalization and ancillary measures for the optim al score level fusion of fingerprint and voice biometrics,

S. T. Mohammed Anzar and P. S. Sathidevi, “On combining multi- normalization and ancillary measures for the optim al score level fusion of fingerprint and voice biometrics,” EURASIP Journal on Advances in Signal Processing, vol. 2014, pp. 1–17, 2014

work page 2014
[3]

An efficient pso optimized integration weight estimation using d -prime statistics for a mu ltibiometric system,

S. Anzar and P. Sathidevi, “An efficient pso optimized integration weight estimation using d -prime statistics for a mu ltibiometric system,” International Journal on Bioinformatics and Biosciences, vol. 2, no. 3, pp. 1504–1511, 2012

work page 2012
[4]

Touchless fingerprint recognition system,

P. P. Navya Sheregar, Vrutti Mistry, “Touchless fingerprint recognition system,” International Journal of Computational Engineering Research (IJCER), April 2021

work page 2021
[5]

Mobile touchless fingerprint ac-quisition and enhancement system,

B. A. Seref Sagiroglu, Mehtap Ulker, “Mobile touchless fingerprint ac-quisition and enhancement system,” in IEEE Congress on Evolutionary Computation (CEC), 2020

work page 2020
[6]

C2cl: Contact to contactless fingerprint matching,

E. L. A. K. J. Steven A. Grosz, Joshua J. Engelsma, “C2cl: Contact to contactless fingerprint matching,” IEEE Transactions on Information Forensics and Security, vol. 17, 2022

work page 2022
[7]

A novel pore extraction method for heterogeneous fingerprint images using convolutional neural networks,

E. M. V. P. F. S. Ruggero Donida Labati, Angelo Genovese, “A novel pore extraction method for heterogeneous fingerprint images using convolutional neural networks,” Pattern Recognition Letters, 2017

work page 2017
[8]

Transfer learning using vgg -16 with deep convolutional neural network for classifying images,

S. Tammina, “Transfer learning using vgg -16 with deep convolutional neural network for classifying images,” International Journal of Scien-tific and Research Publications, vol. 9, 2019

work page 2019

Show all 12 references

[9]

A cnn-based framework for comparison of con- tactless to contact -based fingerprints,

A. K. Chenhao Lin, “A cnn-based framework for comparison of con- tactless to contact -based fingerprints,” IEEE Transactions on Information Forensics and Security, vol. 14, 2019

2019
[10]

An overview of touchless 2d fingerpri nt recognition,

N. B. C. B. M. M. Jannis Priesnitz, Christian Rathgeb, “An overview of touchless 2d fingerpri nt recognition,” EURASIP Journal on Image and Video Processing, 2021

2021
[11]

Mosaicing touchless and mirror-reflected fingerprint images,

J. K. Heeseung Choi, Kyoungtaek Choi, “Mosaicing touchless and mirror-reflected fingerprint images,” IEEE Transactions on Information Forensics and Security, vol. 5, March 2010

2010
[12]

Full 3d touchless fingerprint recognition: Sensor, database and baseline performance,

L. B. Javier Galbally, Gunnar Bostrom, “Full 3d touchless fingerprint recognition: Sensor, database and baseline performance,” in IEEE Inter-national Joint Conference on Biometrics (IJCB), 2017. 6

2017

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Optimal score level fusion using modalities reliability and separability measures,

S. Anzar and P. Sathide vi, “Optimal score level fusion using modalities reliability and separability measures,” Int. J. Comput. Appl, vol. 51, no. 16, pp. 1–8, 2012

work page 2012

[2] [2]

On combining multi- normalization and ancillary measures for the optim al score level fusion of fingerprint and voice biometrics,

S. T. Mohammed Anzar and P. S. Sathidevi, “On combining multi- normalization and ancillary measures for the optim al score level fusion of fingerprint and voice biometrics,” EURASIP Journal on Advances in Signal Processing, vol. 2014, pp. 1–17, 2014

work page 2014

[3] [3]

An efficient pso optimized integration weight estimation using d -prime statistics for a mu ltibiometric system,

S. Anzar and P. Sathidevi, “An efficient pso optimized integration weight estimation using d -prime statistics for a mu ltibiometric system,” International Journal on Bioinformatics and Biosciences, vol. 2, no. 3, pp. 1504–1511, 2012

work page 2012

[4] [4]

Touchless fingerprint recognition system,

P. P. Navya Sheregar, Vrutti Mistry, “Touchless fingerprint recognition system,” International Journal of Computational Engineering Research (IJCER), April 2021

work page 2021

[5] [5]

Mobile touchless fingerprint ac-quisition and enhancement system,

B. A. Seref Sagiroglu, Mehtap Ulker, “Mobile touchless fingerprint ac-quisition and enhancement system,” in IEEE Congress on Evolutionary Computation (CEC), 2020

work page 2020

[6] [6]

C2cl: Contact to contactless fingerprint matching,

E. L. A. K. J. Steven A. Grosz, Joshua J. Engelsma, “C2cl: Contact to contactless fingerprint matching,” IEEE Transactions on Information Forensics and Security, vol. 17, 2022

work page 2022

[7] [7]

A novel pore extraction method for heterogeneous fingerprint images using convolutional neural networks,

E. M. V. P. F. S. Ruggero Donida Labati, Angelo Genovese, “A novel pore extraction method for heterogeneous fingerprint images using convolutional neural networks,” Pattern Recognition Letters, 2017

work page 2017

[8] [8]

Transfer learning using vgg -16 with deep convolutional neural network for classifying images,

S. Tammina, “Transfer learning using vgg -16 with deep convolutional neural network for classifying images,” International Journal of Scien-tific and Research Publications, vol. 9, 2019

work page 2019