{"id":"056f46ad-7dca-4aea-830f-f7942e6d533b","arxiv_id":"2501.05034","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A self-supervised segmentation model detects synthetic fingerprint mosaicking artifacts and a new score quantifies their severity, but real-artifact validation is missing.","lead":"This paper trains a neural network to detect stitching errors in fingerprint images by creating synthetic errors and learning to spot them. It also introduces a scoring metric for artifact severity and shows that stitching artifacts can roughly double fingerprint-matching error rates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The trained model's detection of real mosaicking errors is unvalidated: all ground truth comes from the same synthetic patch/line offset pipeline used for training, so the out-of-distribution results do not test real artifacts.","rationale":"The reader and I identify the same load-bearing concern: the synthetic artifact distribution is the only source of ground truth, so the strong claim about detecting real mosaicking errors is not tested. I find no additional internal inconsistency or methodological error serious enough to change the condition: the architecture and training are standard, the experiments on synthetic data are internally consistent, and the paper's explicit future-work statement about real artifacts corroborates the validation gap. The proposed held-out realistic-mosaicking simulation provides a decisive, feasible check while avoiding the cost of expert annotation. Hence the verdict remains conditional on such validation.","tokens_in":15512,"tokens_out":3221,"duration_ms":32178,"concrete_test":"Build a held-out set of realistic mosaicking errors using an independent simulator: take two overlapping partial fingerprint impressions, register them with a deliberately wrong transform (e.g., a small rotational or translational error), and fuse them with alpha blending along the seam to produce a composite image with a known misalignment mask. Run the trained CL and PR models on this set and measure segmentation IoU/F1 against the known mask. If IoU falls substantially below the Table II values (e.g., below 0.8), the synthetic proxy assumption is invalid and the central claim of detecting real mosaicking artifacts fails; if IoU remains comparably high, the proxy is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the model 'effectively identifies mosaicking errors' on rolled, pressed, and contactless fingerprints rests entirely on synthetic artifacts generated by the same patch/line-offset procedure described in §II.B.3 that was used to create training labels. Table II reports IoU/F1 against these synthetic masks for every dataset, including NIST 300a rolled and slap images, so the evaluation is circular with the training signal: it measures the model's ability to reproduce the augmentation pipeline, not its ability to detect real stitching seams, blended discontinuities, or non-rigid misalignments produced by actual mosaicking algorithms. The paper itself reserves 'supervised finetuning ... on real fingerprint mosaicking artifacts' and 'extension ... to real mosaicking artifact data' for future work (§IV.F, §V), acknowledging that no real ground truth has been used. Consequently, the robustness results and EER analysis (§III.B, §III.C) also pertain only to synthetic offset artifacts; they say nothing about whether the model would flag a genuinely mis-stitched rolled fingerprint. Without a test against independently generated realistic mosaicking errors or expert-annotated real artifacts, the out-of-distribution results in Table II cannot support the strong generalization claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a self-supervised deep learning method for detecting and scoring hard mosaicking artifacts in fingerprint images. A UNet++/ResNeSt-50d segmentation model is trained on large unlabeled single-shot contactless and pressed fingerprint datasets, with supervisory labels generated by a synthetic augmentation pipeline that injects patch offsets and line shifts. The authors also introduce a mosaicking artifact score S and evaluate the model on in-distribution data, on NIST 300a rolled and slap images, on pressed and rolled datasets from a later acquisition, on synthetic fingerprints with simulated skin alterations, and through an EER study of three fingerprint matchers.","tokens_in":15787,"tokens_out":2505,"duration_ms":27973,"significance":"If the proposed model indeed detects real mosaicking artifacts across modalities and sensors it was not trained on, this would be a practically valuable quality-control tool for rolled and contactless fingerprint acquisition, where artifacts can shift minutiae and degrade recognition accuracy. The strengths of the paper include the use of large real-world training sets, a self-supervised pipeline that avoids manual annotation, out-of-distribution evaluation on NIST 300a and subsequent acquisition batches, robustness tests against synthetic skin defects, and a multi-ABIS EER analysis. The central weakness is that every quantitative detection result is measured against ground-truth masks generated by the same synthetic artifact injection procedure that produces the training signal, so the reported IoU/F1 figures do not yet demonstrate detection of real mosaicking errors.","major_comments":[{"comment":"The evaluation of model performance on NIST 300a rolled and slap images, as well as on PRD-1-Test, PRD-2, and ROD-1, uses ground-truth masks created by the same patch-offset and line-shift augmentation procedure described in Section II.B.3 that generates the training labels. The IoU, F1, F2, and recall values in Table II therefore measure how well the model reproduces the synthetic artifact generator, not how well it detects real stitching seams, alpha-blended discontinuities, or non-rigid misalignments produced by actual mosaicking algorithms. This is a load-bearing point because the abstract and conclusion claim high accuracy on rolled, pressed, and contactless fingerprints. The authors should validate on independently generated realistic mosaicking errors or on expert-annotated real mosaicked images, or alternatively temper the generalization claims to synthetic artifacts only.","section":"II.B.3 and Table II"},{"comment":"The paper itself states that supervised finetuning on real fingerprint mosaicking artifacts and extension to real mosaicking artifact data are future work (Section IV.F, Section V). This directly acknowledges that no real ground-truth artifacts were used anywhere in the evaluation. Consequently, the robustness results in Section III.B and the score distribution analysis in Section III.A, while informative about false positives on clean images, do not establish that the model flags genuinely mis-stitched rolled or pressed fingerprints. I recommend that the manuscript either add a real-data validation study or explicitly restrict the central claims to the synthetic artifact setting.","section":"IV.F and V"},{"comment":"The mosaicking artifact score S depends on hand-chosen weights b=5 and c=0.025, and the detection threshold for real images is identified with the patch weight b. The paper validates this threshold only through synthetic experiments and through the observation that clean images have low scores; it does not calibrate the threshold against real mosaicking artifacts or against images that human experts judge as erroneous. Since the score is proposed as an automated screening tool (Section IV.D), the threshold choice should be justified with real positives and negatives, or presented as a tunable parameter with clear guidance for calibration.","section":"Eq. (1) and IV.D"},{"comment":"The EER analysis demonstrates that synthetic small and large offset artifacts degrade the accuracy of SourceAFIS, Bozorth3, and Innovatrics IDKit. This is a useful result, but it should not be presented as validating the detector's real-world utility: the artifacts in Table IV are generated with the same offset ratios used in training (2% to 7% for large offsets), and the table contains no comparison between images the detector flags and images that are actually contaminated by real mosaicking errors. The connection between the EER increase and the artifact score would be stronger if the EER were measured on images whose artifact scores were confirmed against independent ground truth.","section":"III.C and Table IV"}],"minor_comments":[{"comment":"The caption of Table III refers to the second model as the \"CB Model\", while the text and the rest of the paper call it the \"PR Model\"; please make the terminology consistent.","section":"Table III"},{"comment":"There are typos in Section III.B: \"medium woulds\" should be \"medium wounds\", and in Table III the header uses \"tildelow\" where a tilde symbol is intended. These should be corrected.","section":"III.B"},{"comment":"Equation (1) is typeset with ambiguous summation notation (e.g., \"nX i\") and some variables are not defined before the equation. Please rewrite the equation in standard summation form and define all symbols clearly.","section":"Eq. (1)"},{"comment":"NIST 300a is cited as reference [13] with the title \"NIST special database 300\"; please clarify whether the dataset is Special Database 300 or Special Publication 300a and cite the appropriate publication.","section":"References"},{"comment":"The text repeatedly refers to supplementary materials for score distributions, alteration examples, and loss plots, but the arXiv version does not appear to include them. The authors should either include the supplementary material or state where it is available.","section":"Supplementary materials"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable applied contribution, but its central claim of detecting real mosaicking artifacts is not yet supported because all detection ground truth comes from the same synthetic pipeline used for training. I see this as fixable within the manuscript's scope either by adding a real annotated dataset or an independent realistic artifact simulator, or by substantially narrowing the claims. I would not reject the paper outright, but the current version overstates the evidence for cross-modal generalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper is the first I've seen to take on hard mosaicking artifact detection in rolled and contactless fingerprints, and the self-supervised pipeline is genuinely task-specific: synthesize patch and line offsets, train a standard UNet++/ResNeSt segmenter, and report near-perfect IoU on test sets drawn from the same generator. That part is solid engineering. What the paper does not do is test the model on real mosaicking artifacts. Every evaluation, including NIST 300a rolled/slap and their own ROD-1, uses ground truth masks generated by the same patch/line injection procedure used for training. The authors acknowledge this in IV.F and V, where supervised finetuning on real artifacts is listed as future work, but the abstract and conclusion overstate what's been shown: the model generalizes across sensor modalities for synthetic artifacts, not for actual stitching seams.\n\nWhat's genuinely useful: the artifact score gives a simple severity metric, even if it's hand-weighted; the EER study with three matchers shows that 1-2% offsets already double error rates, which motivates the problem well; and the robustness check against scars, noise, and skin defects is a nice touch. The out-of-distribution modality testing (contactless-trained on rolled ink prints) is a legitimate domain transfer test — for synthetic artifacts.\n\nThe soft spots are real but addressable. The main one is circularity: training labels and test ground truth come from the same generator, so the high IoU numbers partly measure how well the model reproduces the generator's rules. Without an evaluation on real mosaicked fingerprints with expert labels, or at least on artifacts produced by an actual mosaicking algorithm, the central claim is unproven. Second, there is no baseline comparison — no edge/sharpness heuristic, no supervised model on the same synthetic data — so we don't know whether a much simpler detector would match these numbers. Third, the score's detection threshold is the hand-chosen patch weight b=5; the observed band structure at multiples of 5 is an artifact of the score definition, not evidence of natural clustering. The 0.061% false-positive claim is threshold-dependent and unvalidated on real data.\n\nOverall: a well-written, honest-in-places engineering paper with a real gap in the literature, but a load-bearing validation gap. I'd send it to peer review — a good referee can push for real-data validation and baselines. I'd cite it for the synthetic pipeline and problem framing, not for the claim that real mosaicking artifacts are detected.","headline":"Useful, clearly-written self-supervised detector for synthetic fingerprint mosaicking artifacts, but the central claim about real artifacts is untested because all ground truth comes from the same injection pipeline used for training.","tokens_in":16278,"tokens_out":3137,"would_cite":true,"duration_ms":31300,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a self-supervised segmentation model, trained only on fingerprints synthetically corrupted with shifted patches and lines, can detect hard mosaicking artifacts in fingerprint images from sensors and modalities it…","keywords":["fingerprint mosaicking","artifact detection","self-supervised learning","image segmentation","biometric quality","artifact score","out-of-distribution generalization"],"falsifier":"Collect genuine rolled fingerprints that contain known mosaicking seams from a production acquisition system, have human experts annotate the artifact regions, and compute the model's IoU on those real masks; if the IoU falls substantially below the 0.9–0.98 range reported on synthetic artifacts, the central claim fails. A weaker but simpler check is to compare the model's artifact score distribution on real mosaicked images against expert severity ratings and see whether the threshold of 5 separates acceptable from unacceptable images.","tokens_in":15344,"feed_emoji":"🖐️","tokens_out":8176,"duration_ms":67407,"temperature":0.7,"pith_summary":"This paper tries to establish that a deep segmentation model can detect hard mosaicking artifacts in fingerprint images without any manually labeled real examples, using only self-supervised training on images the authors synthetically corrupt with shifted patches and lines. The central claim is that this suffices for the model to localize artifacts in fingerprint modalities and sensors it never saw during training, including rolled, slap, and pressed fingerprints, and to remain stable under skin defects, scars, and noise. If true, this would give biometric systems an automated quality check that catches stitching errors before they can shift minutiae and double the equal-error rate of recognition algorithms. The paper also introduces a scalar mosaicking artifact score to quantify severity.","feed_headline":"Self-supervised model finds stitching artifacts in new fingerprints","feed_subtitle":"Trained only on synthetic patch and line offsets, it generalizes to rolled, pressed, and contactless prints.","key_machinery":"The load-bearing mechanism is the self-supervised artifact-generation pipeline, which creates input–mask pairs by selecting up to four patches (each 5–15% of the image dimensions) and offsetting them by 2–7% of the image size, or with 25% probability shifting pixels along a full vertical or horizontal line. A U-Net-style segmentation network with an encoder pretrained on a large general image dataset is trained with Jaccard (intersection-over-union) loss to reproduce these masks. The second component is the mosaicking artifact score: a weighted sum over detected patch areas and line widths, normalized by image area, whose detection threshold is the patch weight $b = 5$, meaning a score at or above 5 indicates at least one closed patch-sized misalignment.","core_discovery":"The discovery is that synthetic patch and line offsets, generated at training time, form a sufficient supervisory signal for a U-Net-style segmentation network to learn the concept of a hard mosaicking artifact. On the contactless test set the model reaches an IoU of 0.982, and on out-of-distribution rolled and slap fingerprints from a public benchmark it retains IoUs of 0.908 and 0.959, with accuracy near 1.000. The same model trained on only 26k pressed fingerprints performs similarly on a later pressed dataset and on rolled fingerprints from the same sensor line. The paper further reports that small stitching offsets raise the equal-error rate of three fingerprint identification systems by 112–132%, which is the practical motivation for the detector.","pith_inferences":["The synthetic proxy leaves untested the hardest real cases, such as alpha-blended seams or elastic deformations from finger rolling; a small annotated set of real mosaicked fingerprints would reveal how much performance drops.","The score's weighting (patch weight 5, line weight 0.025) is a design choice; calibrating it against equal-error rate or human ratings on real data could change the threshold and improve its usefulness as a quality metric.","The reported false-match rate of 0.061% was measured on a test set that itself contains synthetic artifacts; the false-alarm rate on databases with genuine mosaicking errors and rare artifacts could differ.","A natural next step is to apply the score to historical fingerprint databases to estimate how often severe mosaicking artifacts occur in practice."],"forward_implications":["The detector can be deployed as an automated screen for rolled fingerprints, a modality that currently lacks a dedicated quality metric, flagging images for re-acquisition before they enter central registers.","Because the model transfers across modalities with IoU above 0.9, a single training run on one unlabeled corpus can serve sensors of other types.","The mosaicking artifact score, with its threshold at the patch weight of 5, gives a portable severity measure suitable for large-scale database screening; on the pressed test set the false-match rate is 0.061%.","Removing or avoiding images with detected artifacts should reduce the 112–132% equal-error-rate increases that small and large offsets cause in three tested identification systems.","Since training needs no manual annotations, the pipeline can be retrained on new unlabeled fingerprint repositories as they become available."],"supporting_citations":[{"why":"Supplies the 245k-image contactless single-shot corpus used for self-supervised training and testing.","marker":"[36]"},{"why":"Provides the public rolled and slap fingerprint benchmark used to test out-of-distribution generalization.","marker":"[13]"},{"why":"Supplies synthetic fingerprint alteration types (scars, wounds, noise, ink) used to test robustness.","marker":"[25]"},{"why":"Defines the U-Net encoder–decoder structure that the segmentation model builds on.","marker":"[26]"},{"why":"Provides the segmentation-models implementation used to assemble the UNet++ architecture.","marker":"[14]"},{"why":"FingerNet segmentation is used in one of the three ABIS pipelines for equal-error-rate measurement.","marker":"[31]"},{"why":"SourceAFIS matcher is paired with FingerNet for one ABIS pipeline in the equal-error-rate experiments.","marker":"[32]"}],"fun_headline_variants":["Self-supervised AI spots fingerprint stitching errors","No labels needed: AI spots fingerprint mosaicking errors","Self-supervised model scores fingerprint stitching flaws","One model catches stitching artifacts across fingerprint types","Self-supervised AI flags stitching flaws that degrade fingerprint IDs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that synthetic patch and line offsets, produced by the same augmentation pipeline used for training, are a faithful proxy for real mosaicking artifacts in rolled, pressed, and contactless fingerprints; this proxy is never validated against manually annotated real mosaicked images.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised AI spots fingerprint stitching errors","No labels needed: AI spots fingerprint mosaicking errors","Self-supervised model scores fingerprint stitching flaws","One model catches stitching artifacts across fingerprint types","Self-supervised AI flags stitching flaws that degrade fingerprint IDs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000926,"raw_usage":{"total_tokens":3920,"prompt_tokens":849,"completion_tokens":3071,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":3000}},"tokens_in":465,"tokens_out":3071,"duration_ms":20473,"temperature":1.0,"reasoning_tokens":3000,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:20:05.065950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect genuine rolled fingerprints that contain known mosaicking seams from a production acquisition system, have human experts annotate the artifact regions, and compute the model's IoU on those real masks; if the IoU falls substantially below the 0.9–0.98 range reported on synthetic artifacts, the central claim fails. A weaker but simpler check is to compare the model's artifact score distribution on real mosaicked images against expert severity ratings and see whether the threshold of 5 separates acceptable from unacceptable images.","supporting_citations":[{"cited_title":"SourceAFIS fingerprint matcher","cited_arxiv_id":null,"evidence_quote":"Supplies the 245k-image contactless single-shot corpus used for self-supervised training and testing."},{"cited_title":"NIST special database 300: uncompressed plain and rolled images from fingerprint cards","cited_arxiv_id":null,"evidence_quote":"Provides the public rolled and slap fingerprint benchmark used to test out-of-distribution generalization."},{"cited_title":"SynCoLFinGer: Synthetic contactless finger- print generator","cited_arxiv_id":null,"evidence_quote":"Supplies synthetic fingerprint alteration types (scars, wounds, noise, ink) used to test robustness."},{"cited_title":"U-Net: Convo- lutional Networks for Biomedical Image Segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the U-Net encoder–decoder structure that the segmentation model builds on."},{"cited_title":"Segmentation Models Pytorch","cited_arxiv_id":null,"evidence_quote":"Provides the segmentation-models implementation used to assemble the UNet++ architecture."},{"cited_title":"FingerNet: An unified deep network for fingerprint minutiae extraction","cited_arxiv_id":null,"evidence_quote":"FingerNet segmentation is used in one of the three ABIS pipelines for equal-error-rate measurement."},{"cited_title":"Fingerprint Processing: Running the NIST NBIS Fingerprint Toolset on an ODROID-XU4","cited_arxiv_id":null,"evidence_quote":"SourceAFIS matcher is paired with FingerNet for one ABIS pipeline in the equal-error-rate experiments."}],"review_version":1}