{"id":"476e0ad2-b989-433b-b8e6-6a43013cc26f","arxiv_id":"2502.04680","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Image enhancement combined with transfer learning improves touchless fingerprint classification accuracy for VGG-16, VGG-19, and ResNet-50 on the IIT Bombay database, but the reported protocol has gaps.","lead":"This paper compares four pretrained image classifiers, with and without an image enhancement pipeline, on a public touchless fingerprint dataset. It reports that VGG-16 with enhancement reaches higher accuracy than without, suggesting preprocessing helps touchless fingerprint recognition.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Preprocessing gain is confounded: the two arms use different train/test sizes (3200/1600 vs 1800/600) and no subject-disjoint split is stated, so leakage or dataset-size effects could explain the reported improvements.","rationale":"The reader's weakest assumption is the same one I would stress: the comparison between the two pipelines is not controlled. The two arms differ in dataset sizes and possibly in split semantics, so the central causal claim is not currently supported by the reported numbers. I agree with the CONDITIONAL verdict. There is no internal contradiction that forces rejection; the directional consistency across models is mildly supportive, but the confounds mean the specific accuracy values are not trustworthy. A matched, subject-disjoint rerun is feasible with the existing 800-image dataset and would settle whether preprocessing is the cause. No code, preprocessing hyperparameters, or repeated-run statistics are supplied, so external verification is currently impossible. This is an addressable reporting and experimental-design gap, which supports CONDITIONAL rather than REJECT or ACCEPT.","tokens_in":5634,"tokens_out":4743,"duration_ms":48643,"concrete_test":"Ask the authors to release a subject-ID-indexed split and rerun both conditions on the same subject-disjoint split with matched training and test sizes (e.g., 1,800 training and 600 test images per condition, augmentation applied only to the training partition). If preprocessed VGG-16 still reaches 98% test accuracy while the direct method stays at 93% under this matched protocol, the central claim is supported; if the gap collapses or reverses, the reported gain was an artifact of leakage or dataset-size differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The causal claim that preprocessing improves transfer-learning accuracy requires the two compared pipelines to differ only in preprocessing. Section V.A shows they do not: the direct arm has 3,200 training and 1,600 test images, while the preprocessed arm has 1,800 training and 600 test images. Thus training-set size, test-set size, and augmentation ratio are all confounded with the preprocessing variable. The paper also never states whether the train/test split is subject-disjoint. Since the base dataset has 200 subjects with four touchless samples each, an image-level split can place the same subject in both training and testing; if augmented or preprocessed versions of the same original image cross the split, the model can memorize identities rather than learn generalizable features. This would inflate accuracy and could explain the pattern in Tables I and II, including the suspiciously large jumps for the weakest models (ResNet-50 from 0.64 to 0.76, Inception-V3 from 0.17 to 0.64). The reported metrics come from a single run with no error bars, no code, and no preprocessing parameter values, so the 93%-to-98% gap is not established as statistically significant. The concern is not that preprocessing cannot help; it is that the experiment as described cannot distinguish preprocessing from leakage or dataset-size effects.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":5866,"tokens_out":4203,"duration_ms":39391,"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":[{"comment":"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":"Section V.A"},{"comment":"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.","section":"Section V.A"},{"comment":"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":"Abstract, Section V.C"},{"comment":"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.","section":"Section V.B"}],"minor_comments":[{"comment":"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.","section":"Section IV"},{"comment":"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.","section":"Keywords"},{"comment":"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.","section":"Tables I and II"},{"comment":"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.","section":"Figures 3 and 4"},{"comment":"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.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the qualitative pattern is plausible, but the confounded experimental design and the abstract/table inconsistency are load-bearing issues. I recommend major revision rather than rejection because a cleaner comparison (matched dataset sizes, subject-disjoint split, repeated runs) is achievable within the manuscript's scope. The absence of code or preprocessing parameters further limits verification, so I would encourage the authors to provide these as supplementary material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.","headline":"A standard benchmark whose central comparison is confounded by different dataset sizes and an unspecified split; the preprocessing gain is plausible but not established.","tokens_in":6422,"tokens_out":2453,"would_cite":false,"duration_ms":25288,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["touchless fingerprint recognition","transfer learning","image enhancement","VGG-16","CLAHE","SIFT","biometric authentication","IIT-Bombay fingerprint database"],"falsifier":"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.","tokens_in":5406,"feed_emoji":"🖐️","tokens_out":8393,"duration_ms":68924,"temperature":0.7,"pith_summary":"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.","feed_headline":"Preprocessing lifts touchless fingerprint accuracy to 98 percent","feed_subtitle":"In the paper's tests, every one of four pretrained networks improves after the same enhancement chain.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Motivates preprocessing as essential for feature extraction in touchless fingerprint recognition.","marker":"[4]"},{"why":"Shows a mobile touchless fingerprint system with an integrated enhancement pipeline improves recognition.","marker":"[5]"},{"why":"Highlights consistent image quality through preprocessing for matching contact-based and contactless fingerprints.","marker":"[6]"},{"why":"Provides the VGG-16 transfer-learning baseline the paper builds on for touchless fingerprint classification.","marker":"[8]"},{"why":"Supports the role of preprocessing in aligning features between contactless and contact-based fingerprints.","marker":"[9]"},{"why":"Reviews touchless 2D fingerprint recognition and stresses advanced preprocessing for reliable systems.","marker":"[10]"}],"fun_headline_variants":["Enhancement boosts touchless fingerprint accuracy across four CNNs","Preprocessing aids touchless fingerprint CNNs across the board","Image enhancement before transfer learning lifts touchless print accuracy","Indirect beats direct: enhancement improves touchless fingerprint recognition","Touchless fingerprint CNNs all improve with image enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Enhancement boosts touchless fingerprint accuracy across four CNNs","Preprocessing aids touchless fingerprint CNNs across the board","Image enhancement before transfer learning lifts touchless print accuracy","Indirect beats direct: enhancement improves touchless fingerprint recognition","Touchless fingerprint CNNs all improve with image enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001737,"raw_usage":{"total_tokens":6842,"prompt_tokens":903,"completion_tokens":5939,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":5858}},"tokens_in":519,"tokens_out":5939,"duration_ms":45150,"temperature":1.0,"reasoning_tokens":5858,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:52:00.894901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Touchless fingerprint recognition system,","cited_arxiv_id":null,"evidence_quote":"Motivates preprocessing as essential for feature extraction in touchless fingerprint recognition."},{"cited_title":"Mobile touchless fingerprint ac-quisition and enhancement system,","cited_arxiv_id":null,"evidence_quote":"Shows a mobile touchless fingerprint system with an integrated enhancement pipeline improves recognition."},{"cited_title":"C2cl: Contact to contactless fingerprint matching,","cited_arxiv_id":null,"evidence_quote":"Highlights consistent image quality through preprocessing for matching contact-based and contactless fingerprints."},{"cited_title":"Transfer learning using vgg -16 with deep convolutional neural network for classifying images,","cited_arxiv_id":null,"evidence_quote":"Provides the VGG-16 transfer-learning baseline the paper builds on for touchless fingerprint classification."},{"cited_title":"A cnn-based framework for comparison of con- tactless to contact -based fingerprints,","cited_arxiv_id":null,"evidence_quote":"Supports the role of preprocessing in aligning features between contactless and contact-based fingerprints."},{"cited_title":"An overview of touchless 2d fingerpri nt recognition,","cited_arxiv_id":null,"evidence_quote":"Reviews touchless 2D fingerprint recognition and stresses advanced preprocessing for reliable systems."}],"review_version":1}