Towards Fingerprint Mosaicking Artifact Detection: A Self-Supervised Deep Learning Approach

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses an important and underexplored problem: detecting hard mosaicking artifacts in fingerprint images without requiring manually annotated artifact data. The proposed self-supervised training strategy, the use of multiple fingerprint modalities, and the attempt to connect artifact detection with matching performance are valuable aspects of the work. The topic is relevant for biometric quality control and has practical importance.

However, the paper needs substantial strengthening before it can be considered ready for publication in a high quality journal. The main concern is that the model is trained and largely evaluated using artificially generated artifacts. This is a reasonable starting point, but the manuscript does not yet provide enough evidence that the simulated patch and line artifacts adequately represent real mosaicking failures produced by actual acquisition devices and stitching algorithms. Some validation on manually verified real mosaicking artifacts, even on a limited subset, would considerably improve the credibility of the claims.

The proposed mosaicking artifact score is useful, but its constants and detection threshold appear empirically selected. The authors should justify these choices more rigorously, for example through sensitivity analysis or threshold calibration using validation data. The current results also need stronger comparison against relevant baselines, such as classical image quality measures, simple edge or discontinuity detectors, or alternative segmentation architectures. Without such comparisons, it is difficult to judge how much improvement is due to the proposed framework rather than the general capacity of a large segmentation model.

The experimental reporting should also be improved. Accuracy is reported as 1.000 in several cases, which may be misleading for highly imbalanced segmentation masks. More emphasis should be placed on precision, false positive rate, false negative rate, confidence intervals, and per-artifact-type performance. The EER analysis is interesting, but the subset size and statistical uncertainty should be reported more clearly. Finally, since the study uses biometric data, the statements on ethics, consent, and privacy restrictions should be clarified. “Not applicable” is not sufficiently informative for real fingerprint datasets.

Overall, the manuscript is promising and potentially publishable, but it requires major revision, particularly in real-artifact validation, baseline comparison, threshold justification, and ethical/data governance clarification.

Comments on the Quality of English Language

The English is generally understandable, but it should be revised for clarity, grammar, and smoother scientific expression. Several sentences are awkwardly phrased, and some terms are used inconsistently, such as “contact-less” and “contactless,” or “mosaick artifact score” and “mosaicking artifact score.” The manuscript would benefit from careful proofreading by a fluent technical editor to improve readability, sentence structure, and consistency throughout the paper.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. The paper directly adopts the existing ResNeSt-50d as the encoder and UNet++ as the decoder, without any targeted improvements or optimizations to the network architecture. It lacks original structural design at the model level, resulting in weak innovation of the proposed algorithm.

2. The study performs self-supervised training by adopting manually synthesized fingerprint mosaicking artifacts. Nevertheless, the similarity between the synthetic artifacts and real-scene mosaicking artifacts is not assessed, and it cannot be proven that synthetic samples can replace real artifacts for model training. It is recommended to add a small number of fingerprint images with real artifact annotations and set up comparative experiments for further verification.

3. The definitions and physical implications of each variable in the mosaicking artifact score formula are ambiguously expressed and insufficiently clarified. It is suggested to supplement detailed explanations for all variables and the internal calculation logic of the formula.

4. The hyperparameters b and c in the mosaicking artifact score formula are directly set to fixed values without stating the selection basis and reasons. It is necessary to supplement comparative experiments with different parameter combinations to demonstrate the rationality and optimality of the current values of b and c.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This revised manuscript presents a technically sound and practically relevant framework for detecting fingerprint mosaicking artifacts using a self-supervised deep learning approach. The topic is important for biometric quality assurance, especially for rolled and contactless fingerprint acquisition pipelines where stitching failures can negatively affect matching reliability.

The revised version is substantially improved compared with the previous submission. In particular, the manuscript now provides a clearer explanation of the synthetic artifact generation strategy, a more rigorous interpretation of the mosaicking artifact score, additional statistical clarification for the EER experiments, and a much more balanced limitations discussion. The inclusion of a real mosaicking failure example also improves the connection between the synthetic training setup and operational failure modes.

The cross-modality evaluation and robustness experiments are valuable additions and demonstrate that the framework generalizes reasonably well across acquisition conditions. The analysis of EER degradation further strengthens the practical relevance of the work.

Despite these improvements, several limitations remain.

First, the framework is still evaluated primarily using synthetically generated artifacts. Although the manuscript now acknowledges this limitation explicitly, the lack of a quantitatively annotated real-world mosaicking artifact benchmark remains the main weakness of the study. The current evidence suggests promising generalization, but the conclusions regarding operational deployment should be stated more cautiously.

Second, the proposed mosaicking artifact score is manually designed using heuristic parameters. While the revised manuscript now explains the rationale behind the parameter choices, the score is not calibrated against expert annotations or operational severity labels. Future work should investigate data-driven calibration or learning-based scoring approaches.

Third, the architectural novelty is relatively limited because the framework relies on existing segmentation architectures. The contribution therefore lies primarily in the self-supervised formulation, artifact simulation strategy, and evaluation pipeline rather than in methodological innovation at the network level. This should continue to be stated clearly to avoid overstating novelty.

There are also several presentation-related issues that should still be improved before publication:

• Some sentences remain overly long and occasionally repetitive, especially in the Discussion section.
• Terminology usage should be made more consistent throughout the manuscript. Terms such as “contact-less/contactless” and “mosaick/mosaicking” are not always used consistently.
• A few claims regarding robustness and deployment readiness should be slightly moderated given the limited real-artifact validation.
• Several figures could benefit from improved readability and slightly larger annotations.

Overall, the paper presents a useful applied contribution and the revised manuscript addresses many of the previous concerns. With minor revision focused mainly on presentation clarity and more cautious interpretation of real-world applicability, the manuscript would be suitable for publication.

Comments on the Quality of English Language

The English has improved compared with the previous version, and the manuscript is generally understandable. However, the paper would still benefit from careful language editing for grammar, sentence structure, consistency of terminology, and scientific readability. Several sections contain overly long sentences and occasional awkward phrasing. Minor inconsistencies in terminology and formatting are also present throughout the manuscript.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Overall, the completeness of this paper is high with distinct core innovations:

First, a manual-annotation-free self-supervised training paradigm is constructed. The paper has supplemented verification that synthetically generated artifacts are highly matched with fingerprint mosaicking artifacts in real scenarios, and the model can effectively identify real artifacts. The design in this part is reasonable and fully demonstrated.

Second, a dedicated quantitative scoring system for artifacts is designed to realize automatic graded evaluation of the severity of fingerprint mosaicking artifacts. Meanwhile, the variable definitions and physical interpretation of the scoring formula have been supplemented and refined, presenting complete logic and strong interpretability.

On this basis, two minor revision suggestions are proposed:

1. Optimize the typesetting of figures in the manuscript. In particular, the order of subfigures in Figure 2 is inconsistent with the description order in the main text, which affects reading comprehension.

2. It is recommended to supplement a set of comparative experiments. If other existing artifact segmentation models in the same field are available, please add them for comparison. If there are no publicly available dedicated models of this kind, adopt the baseline model of original ResNeSt-50d encoder + UNet decoder to conduct comparative experiments with the proposed model. This can more intuitively reflect the advantages of the proposed method and further enhance the persuasiveness of experimental results.

Author Response

Please see the attachment

Author Response File: Author Response.pdf