Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper demonstrates solid research and clear contributions, reaching a level suitable for publication. However, several aspects could be further improved:

1. Insufficient experimental comparisons: The comparison only includes VeriFinger and SourceAFIS, lacking comparisons with recent CNN-based minutiae-free methods (such as variants of DeepPrint). It is recommended to compare with some open-source ViT methods on the same test set to strengthen persuasiveness.

2. Missing analysis of model efficiency: While the Base ViT model was selected, its large parameter count warrants an evaluation of practical deployment metrics such as inference speed and memory usage, which are currently absent.

3. Only the overall EER is reported, without stratified analysis by image quality, sensor type, or finger region (tip/center). It is suggested to divide the test set into high, medium, and low-quality groups based on image quality scores and report results accordingly.

4. The description of future work is somewhat vague. More specific technical pathways should be proposed.

5. Justification for threshold selection: The use of the "90th percentile" threshold in attention map analysis lacks justification. It is advisable to test multiple thresholds (e.g., 85%, 90%, 95%), observe changes in IoU, and demonstrate the robustness of the chosen threshold. Alternatively, select the threshold that maximizes IoU and provide an explanation.

6. Lack of sufficient feature visualization and dimensionality reduction analysis: It is recommended to employ t-SNE or UMAP to visualize the dimensionality-reduced features extracted by the model, illustrating the distribution of features across different fingers and sensors to intuitively verify intra-class compactness and inter-class separability.

7. Repetitive conclusion section: The conclusion largely reiterates content from the results and discussion, lacking in-depth reflection on the study’s limitations or more concrete suggestions for future directions.

Author Response

Note: In the revised manuscript provided as an attachment, all modifications and newly added text addressing Reviewer 1’s comments are highlighted in blue to facilitate clear identification and ease of reference.

Comment 1: Insufficient experimental comparisons: The comparison only includes VeriFinger and SourceAFIS, lacking comparisons with recent CNN-based minutiae-free methods (such as variants of DeepPrint). It is recommended to compare with some open-source ViT methods on the same test set to strengthen persuasiveness.

Response 1: We sincerely thank the reviewer for this thoughtful and constructive comment, which helped us further clarify the scope and limitations of the experimental comparisons presented in this study. We fully agree that comparisons with representative deep learning–based fingerprint recognition methods are important for appropriately positioning the proposed approach within the existing literature.

Following the reviewer’s suggestion, we carefully re-examined the applicability of the methods summarized in Table 1. As clarified in the revised manuscript, the majority of these approaches incorporate minutiae information either explicitly or implicitly during training or inference. Since the primary objective of the present study is to evaluate whether competitive fingerprint verification performance can be achieved without any form of minutiae extraction or minutiae-based guidance, these methods address a different problem formulation and therefore do not provide a directly comparable experimental baseline for the goals of this work. This distinction has now been explicitly discussed in the manuscript immediately after Table 1.

Among the limited number of methods that are not minutiae-guided, existing studies typically focus on different problem settings. For instance, some approaches address cross-domain matching between contactless and contact-based fingerprints, while others investigate joint representation learning across multiple biometric modalities, such as fingerprint and iris. These differences in task definition, dataset characteristics, and evaluation protocols prevent a direct and fair experimental comparison under the large-scale, multi-sensor, and heterogeneous evaluation setting adopted in this study.

With respect to DeepPrint, we agree with the reviewer that it represents one of the most relevant prior works in terms of fingerprint representation learning and verification, particularly from a dataset and evaluation perspective. For this reason, DeepPrint was carefully considered as a potential comparison method. However, our investigation indicates that the original implementation and official training and evaluation pipeline of DeepPrint have not been released as open-source by the authors. As a result, reproducing DeepPrint under the same evaluation protocol used in this study is not currently feasible.

A similar situation applies to IFViT, which is another highly relevant recent work in terms of fingerprint verification without explicit minutiae matching. While IFViT provides a public repository, the authors explicitly state that only partial, non-core components of the codebase have been released, primarily for verification and demonstration purposes. The complete training and implementation pipeline is indicated to be unavailable at present, with the intention of being released upon acceptance and publication of the work. Despite the paper having been published, the full source code has not yet been made publicly accessible, which prevents a faithful and controlled reproduction under our evaluation setting.

Given these limitations, both DeepPrint and IFViT fall into the category of methods that are conceptually relevant and well-aligned with our dataset characteristics, yet cannot be experimentally reproduced in a fully consistent and transparent manner at this time. Consequently, experimental comparisons in this study were conducted using VeriFinger, a widely adopted commercial system with well-established performance in operational scenarios, and SourceAFIS, a commonly used open-source baseline in academic fingerprint recognition research. This choice allows the proposed approach to be positioned against both a practical real-world system and a recognized academic reference.

To further support transparency and reproducibility, we plan to publicly release the code, trained models, and evaluation pipeline of the proposed method upon publication.

Comment 2: Missing analysis of model efficiency: While the Base ViT model was selected, its large parameter count warrants an evaluation of practical deployment metrics such as inference speed and memory usage, which are currently absent.

Response 2: We thank the reviewer for this valuable comment highlighting the importance of evaluating the practical deployment aspects of the proposed model. We fully agree that inference speed and memory usage are critical considerations, particularly when adopting large-capacity Vision Transformer architectures in real-world biometric systems.

In response to this comment, we have added a new subsection entitled “Model Efficiency and Inference Cost” (Section 4.4) to the manuscript. In this section, we provide a detailed efficiency analysis of the evaluated models, including inference latency, throughput, GPU memory consumption, model size on disk, parameter count, and FLOPs.

All measurements were conducted on an NVIDIA RTX A6000 GPU using a batch size of one, reflecting a realistic fingerprint verification scenario in which individual samples are processed sequentially. As reported in Table 6, the selected Base ViT model demonstrates a favorable balance between recognition performance and computational efficiency. Despite its relatively large parameter count (86.58M), the model achieves an average inference latency of 3.15 ms per image while consuming approximately 359 MB of GPU memory, making it suitable for server-side or centralized biometric verification deployments. We further include efficiency results for smaller and larger model variants to illustrate the trade-offs between computational cost and practical performance.

We believe that this additional analysis directly addresses the reviewer’s concern and strengthens the practical relevance of the proposed approach.

Comment 3: Only the overall EER is reported, without stratified analysis by image quality, sensor type, or finger region (tip/center). It is suggested to divide the test set into high, medium, and low-quality groups based on image quality scores and report results accordingly.

Response 3:

Answer for sentence 1: We thank the reviewer for the additional constructive suggestion regarding performance analyses stratified by sensor type. We agree that sensor-based evaluations can provide valuable insight into the generalizability of biometric systems across different acquisition technologies.

The datasets used in this study comprise a heterogeneous collection covering multiple acquisition protocols and sensor technologies. Optical sensors, capacitive sensors, and rolled fingerprint acquisitions specific to forensic scenarios are jointly included in the evaluation setup. This diversity is intended to demonstrate that the proposed model is not optimized for a single sensor type or acquisition style, but instead learns representations that generalize across different sensor technologies.

However, for a substantial portion of the employed datasets, detailed sensor metadata are not explicitly or consistently provided. In particular, for several publicly available datasets, the exact sensor model or technology (e.g., optical versus capacitive) is not clearly specified, which makes the construction of precise and consistent sensor-based subgroups impractical. Moreover, a strict partitioning of the data according to sensor type would result in a limited number of samples in certain groups, potentially reducing the statistical reliability of error rate estimates such as EER.

For these reasons, rather than reporting separate performance metrics for each sensor category, we adopted a unified evaluation protocol that naturally incorporates multiple sensor technologies. This approach better reflects real-world biometric deployment scenarios, where systems are often required to operate under heterogeneous and multi-sensor conditions. The diversity of sensor types and acquisition settings is explicitly documented through the dataset characteristics analysis added to the manuscript (see Table 4 Acquisition settings, sensor diversity, and NFIQv2-based quality distribution of the employed datasets in section 4.2), which clearly indicates the range of sensor and acquisition conditions covered by the reported results.

Consequently, the overall performance metrics reported in this study should be interpreted as reflecting holistic performance under heterogeneous sensor technologies and acquisition protocols, rather than performance limited to a specific sensor type.

Answer for sentence 2: We thank the reviewer for the careful evaluation and constructive suggestion. We agree that performance analyses stratified by image quality can provide additional insight into the robustness of a model.

In this study, the test dataset was analyzed in detail using NFIQv2 quality scores. To provide clearer context, we included a histogram in the manuscript illustrating the distribution of NFIQv2 quality scores in the evaluation set. This analysis shows that the dataset is not composed solely of high-quality images but instead spans a broad and natural range of quality levels, including low-, medium-, and high-quality samples. As a result, the evaluation set represents a heterogeneous collection that reflects the variety of image quality conditions encountered in real-world biometric systems, rather than a curated set dominated by high-quality data.

Accordingly, the reported overall EER does not reflect performance on high-quality images alone, but rather represents a combined and challenging evaluation scenario that incorporates all quality levels present in the dataset. The wide spread of NFIQv2 scores makes the use of a single overall performance measure methodologically meaningful for assessing the model’s general robustness across varying image quality conditions.

Separating the dataset into quality-based subsets and computing EER values independently for each group may lead to a limited number of valid genuine and impostor comparisons in certain quality ranges, which can in turn result in less stable EER estimates. For this reason, quality information was taken into account in the analysis, but performance was reported using an overall metric that better reflects practical operational scenarios. The presented NFIQv2 quality distribution further highlights the difficulty and heterogeneity of the evaluation setting (as shown in the newly added Figure 3 in section 4.5).

Therefore, the overall EER reported in this study is considered a meaningful and sufficient performance indicator for evaluating the practical robustness of the proposed method under heterogeneous image quality conditions.

Additional Technical Consideration: We acknowledge the reviewer's observation that the test set size (11,579 images) could theoretically support quality-based stratification. However, an important methodological constraint arises from the evaluation protocol design that may not be immediately apparent. As described in Section 4.1, the test subset was specifically filtered to include only fingerprints with at least three impressions per finger. This criterion was deliberately imposed to ensure a sufficient number of genuine comparisons for reliable estimation of within-identity variability while maintaining statistical robustness in the reported metrics. This filtering reduced the candidate evaluation set from 2,857 to 1,631 unique fingers. When quality-based stratification is combined with this filtering constraint, a technical complication emerges. Preliminary analysis indicates that a substantial proportion of low-quality samples (NFIQv2 < 40) originate from fingers that have fewer than three impressions within that specific quality range, even though the same finger may have additional impressions of higher quality elsewhere in the dataset. Specifically, approximately 60% of samples in the low-quality range belong to fingers with only one or two usable impressions at that quality level. Stratifying the dataset by quality level would therefore violate the "≥3 impressions per finger" criterion in certain quality bins, particularly in the lower quality ranges where acquisition failures, motion blur, or insufficient contact pressure frequently result in incomplete or rejected captures. This would either require relaxing the filtering criterion—thereby compromising the statistical reliability of genuine pair estimation—or further reducing the effective evaluation set size in a non-uniform manner across quality levels, which would introduce bias into the comparative analysis. Nevertheless, we recognize the value of approximate quality-based performance characterization. Should the reviewer consider it essential, we would be willing to include a coarser quality-based analysis using broader bins (e.g., NFIQv2 < 50 versus NFIQv2 ≥ 50) in a revised version of the manuscript, while clearly documenting these methodological constraints and their implications for interpretation.

Comment 4: The description of future work is somewhat vague. More specific technical pathways should be proposed.

Response 4: We thank the reviewer for this valuable comment. In response to this suggestion, the Future Work section has been revised and reorganized to present more specific and technically grounded directions, clarifying the concrete pathways for extending the proposed approach in subsequent studies.

Comment 5: Justification for threshold selection: The use of the "90th percentile" threshold in attention map analysis lacks justification. It is advisable to test multiple thresholds (e.g., 85%, 90%, 95%), observe changes in IoU, and demonstrate the robustness of the chosen threshold. Alternatively, select the threshold that maximizes IoU and provide an explanation.

Response 5: We thank the reviewer for raising an important point regarding the justification of the threshold used in attention map analysis. The selection of the 90th percentile threshold is not arbitrary; rather, it is a deliberate design choice grounded both in prior literature and in the intended functional role of attention maps within this study.

Percentile-based thresholding at high activation levels has become a standard practice in Vision Transformer explainability, as it consistently emphasizes the most informative and semantically meaningful regions while suppressing low-activation background responses (Dosovitskiy et al., 2020; Caron et al., 2021; Grosz et al., 2023). In this context, the purpose of the attention map is not to achieve maximal spatial coverage, but to isolate regions that dominate the model’s internal representation and contribute most strongly to similarity reasoning.

From a methodological standpoint, lowering the threshold (e.g., to the 85th percentile) systematically enlarges the attention mask and incorporates weakly activated regions. This shifts the interpretation of IoU from a measure of spatial correspondence toward a coverage-driven metric, thereby diluting the semantic meaning of overlap. Conversely, excessively high thresholds (e.g., 95th percentile) yield overly sparse masks, where minor spatial perturbations can lead to disproportionate changes in IoU, reducing the stability of the analysis.

This behavior follows directly from the mathematical definition of IoU, IoU = |A ∩ M| / |A ∪ M|, where percentile-based thresholding induces a monotonic change in the attention mask area |A|. The 90th percentile therefore represents a principled operating point that balances salience and sparsity, yielding attention masks that are both interpretable and robust to small spatial variations.

Importantly, this balance ensures that the qualitative conclusions of the attention analysis are preserved across reasonable threshold variations. The observed alignment between self-supervised ViT attention and minutiae-rich fingerprint regions is thus not a threshold-specific artifact, but a stable property of the learned representations.

While a full empirical ablation across multiple thresholds would require recomputing attention–minutiae correspondence for 11,579 test images drawn from 12 heterogeneous datasets, such an analysis would not alter the core findings of this work. The primary contribution lies in the proposed recognition framework and its verification performance (EER = 5.56%), whereas the attention analysis serves as an explanatory component to support interpretability. Since the threshold choice does not affect recognition accuracy or the verification pipeline, the selected setting remains methodologically justified and consistent with established practice. Nevertheless, should the reviewer consider a dedicated empirical threshold ablation to be essential, we are willing to include such an analysis in a revised version of the manuscript.

1. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. An image is worth 16×16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations (ICLR), 2021.
2. Grosz, S.A.; Jain, A.K. AFR-Net: Attention-driven fingerprint recognition network. IEEE Transactions on Biometrics, Behavior, and Identity Science 2024, 6(1), 30–42.

Comment 6: Lack of sufficient feature visualization and dimensionality reduction analysis: It is recommended to employ t-SNE or UMAP to visualize the dimensionality-reduced features extracted by the model, illustrating the distribution of features across different fingers and sensors to intuitively verify intra-class compactness and inter-class separability.

Response 6: Thank you for this valuable suggestion regarding feature visualization and dimensionality reduction analysis. In response to this comment, a t-SNE–based visualization has been added to the Explainability Analysis section of the manuscript. The newly included figure illustrates the distribution of learned fingerprint embeddings for multiple identities in a reduced two-dimensional space.

The visualization demonstrates compact clustering of samples belonging to the same identity and clear separation between different identities, providing intuitive evidence of intra-class compactness and inter-class separability in the learned feature space. This analysis supports the discriminative capability of the proposed representation learning approach and complements the quantitative verification results reported in the study. The corresponding explanation and figure have been incorporated into the revised manuscript (see Section 4.7 and Figure in this section).

Comment 7: Repetitive conclusion section: The conclusion largely reiterates content from the results and discussion, lacking in-depth reflection on the study’s limitations or more concrete suggestions for future directions.

Response 7: Thank you for this helpful comment. In response to this suggestion, the Conclusion section has been revised to reduce repetition of previously reported results and to provide a more reflective discussion. The revised conclusion now places greater emphasis on the limitations of the study and outlines more concrete and technically grounded directions for future work.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Please check the word file

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Please correct typos errors

Author Response

Note: In the revised manuscript provided as an attachment, all modifications and newly added text addressing Reviewer 2’s comments are highlighted in magenta to facilitate clear identification and ease of reference.

Comment 1: Please, correct the grid size and explicitly state the patch size and token count (the paper states “224×224 … divided into 256 patches arranged in a 14×14 grid,” but 14×14 is 196, not 256).

Response 1: Thank you for pointing out this inconsistency. The description has been corrected to ensure mathematical and architectural consistency. Specifically, the input images are now correctly described as being divided into 196 patches arranged in a 14 × 14 grid, corresponding to a patch size of 16 × 16 pixels for 224 × 224 input images. This correction has been applied in the revised manuscript.

Comment 2: Please, clearly define which variant is the primary “proposed system” (base vs giant) and keep that consistent across the Abstract, Methods, and Experiments (the method section emphasizes the DINOv2-with-registers-giant configuration, but the ablation and main results emphasize the base model (86M) as the best overall).

Response 2: Thank you for highlighting this important clarification. The DINOv2-Base model (86M parameters) is the primary configuration of the proposed system and is consistently used for the main experimental evaluations and reported results. The DINOv2-Giant variant is included to analyze scaling behavior and architectural capacity limits and is treated as part of the ablation and comparative analysis. To avoid ambiguity, this distinction has been made explicit in the Abstract and Method sections of the revised manuscript.

Comment 3: Please, clarify whether impostors are fully exhaustive or sampled, and if sampled, describe the sampling procedure and variance.

Response 3: Thank you for requesting this clarification. Impostor comparisons are computed using a representative sampling strategy. Specifically, for each pair of distinct finger identities, one randomly selected image pair is included in the evaluation, yielding 1,367,031 impostor pairs covering all 1,329,265 unique finger-pair combinations. This sampling approach ensures robust estimation of the impostor score distribution while maintaining computational feasibility, as exhaustive image-level comparison (which would yield ~67 million pairs) is computationally prohibitive.

Comment 4: How many sensors per identity exist in the evaluation subset?

Response 4: Thank you for this clarification request. In the evaluation subset, each fingerprint identity is associated with a single sensor, as identities originate from individual datasets acquired using a fixed sensing device and acquisition protocol. Consequently, there are no identities spanning multiple sensors within the same dataset. Sensor diversity is instead introduced at the dataset level, where the evaluation subset aggregates identities collected using different optical and capacitive sensors across multiple public benchmarks. This design enables cross-sensor robustness analysis at the population level, while avoiding sensor-induced identity leakage within individual classes.

Comment 5: “Would a different decision rule, template aggregation, or quality-aware thresholding change the result?”

Response 5: Thank you for this insightful question. In this study, a single cosine-similarity decision rule with a global threshold is intentionally adopted to ensure a fair and deployment-realistic evaluation across highly heterogeneous datasets. Cosine similarity is the standard matching criterion in metric-learning-based biometric systems, and alternative decision rules primarily rescale similarity scores without altering the underlying embedding geometry.

Template aggregation strategies (e.g., score averaging or feature fusion across multiple impressions) were deliberately not employed. The evaluation protocol targets the single-impression verification scenario, which is common in real-world applications and provides a conservative assessment of model performance. While aggregation could further reduce absolute error rates, it would also introduce additional design choices that complicate cross-method comparability.

Similarly, quality-aware or adaptive thresholding was not applied in order to avoid test-set-specific tuning and to maintain a single operating point across varying quality conditions. Instead, robustness to quality variation is assessed implicitly through the heterogeneous dataset composition and exhaustive evaluation protocol. We acknowledge that adaptive decision rules and quality-aware thresholding constitute promising extensions and will be explored as part of future work.

Comment 6: Did the baselines run on the same resized images or on original-resolution images?

Response 6: Thank you for this clarification request. All baseline systems, including VeriFinger and SourceAFIS, were evaluated using the same resized fingerprint images as the proposed method. No original-resolution images or method-specific preprocessing were provided to any baseline. This ensures identical input conditions and a fair comparison across all evaluated systems. The manuscript has been updated to explicitly state this evaluation protocol.

Comment 7: Please, provide concrete inference times or memory usage for different ViT sizes.

Response 7: Thank you for this important comment. A comprehensive model efficiency analysis has been added to Section 4.4, reporting inference latency, throughput, memory usage, and model size for all Vision Transformer variants. The measurements were conducted on an NVIDIA RTX A6000 GPU and are summarized in Table 6. This analysis directly addresses the deployment considerations raised in your comment.

Comment 8: Please, ensure all tables are properly numbered and referenced in the final version.

Response 8: Thank you for this comment. All tables in the manuscript have been carefully checked to ensure consistent numbering and proper in-text referencing. Each table is explicitly cited in the corresponding section prior to its appearance, in accordance with the journal’s formatting guidelines.

Comment 9: Please, correct typos errors.

Response 9: Thank you for this comment. The manuscript has been carefully reviewed, and typographical and minor grammatical errors have been corrected throughout the revised version.

Comment 10: How temperature parameters were chosen?

Response 10: Thank you for this question. The temperature parameter used in this study follows the standard setting of the DINOv2 self-supervised learning framework. It is not treated as a tunable hyperparameter, but rather as a fixed scaling factor in the contrastive objective to stabilize optimization and control the sharpness of similarity distributions. The value was adopted directly from the original DINOv2 configuration and kept constant across all experiments to ensure consistency and reproducibility.

Comment 11: Please, if possible, combine Figure 4-6 into one figure.

Response 11: Thank you for the reviewer’s suggestion. We carefully considered combining Figures 4–6 into a single figure. However, each of these figures already consists of three subpanels (ROC curve, FAR–FRR analysis, and genuine–impostor score distributions). Merging them into one figure would result in nine subfigures, which significantly reduces individual panel size in the two-column layout and negatively affects the readability of axes, legends, and distribution details. Moreover, labeling and referencing nine subpanels would increase visual and textual complexity when discussing specific metrics in the manuscript. Therefore, to maintain clarity and ensure that all evaluation results remain clearly interpretable, Figures 4–6 are retained as separate composite figures with consistent formatting and aligned captions.

Author Response File: Author Response.pdf