Finger Type Classification for Fingerprint Image Error Correction in Large Scale Biometric Databases

11 12

- Large-scale biometric systems, essential for national security and border 13 management, increasingly rely on multimodal databases containing millions of identities. 14 However, operational pressures and insufficient training lead to frequent image classification and 15 labeling errors by human operators. These critical data integrity issues include the mislabeling of 16 rolled vs. flat fingerprints, out-of-sequence captures, and the insertion of incorrect modalities. 17 Such errors render enrollment records unreliable, compromising subsequent identity verification 18 processes. Since manually sorting vast image archives is unfeasible, our study proposes an 19 automated solution. The primary objective was to deploy a Siamese Network to classify 20 fingerprints by their precise finger type and collection methodology (flat or rolled impressions). 21 A secondary, but central, goal was to investigate the influence of varying embedding dimensions 22 (64, 128, 256, 512) and similarity thresholds (0.5, 0.2, 0.1) on the network's performance 23 metrics. Our most significant finding demonstrates a clear trade-off: a lower similarity threshold 24 drastically increases conditional accuracy and precision (e.g., up to 98%) but simultaneously 25 increases the proportion of images categorized as "uncertain" (up to 24%). In a practical, large-26 scale application, this necessitates balancing superior classification accuracy against a higher 27 volume of images requiring costly manual inspection. This work provides a proof-of-concept 28 tool capable of efficiently quantifying the percentage of images requiring human review across 29 various modalities (fingerprints, face, iris). The eventual goal is a lightweight, efficient tool to 30 establish standard preprocessing procedures for any large biometric dataset, dramatically 31 reducing the time and cost associated with data integrity maintenance. 32 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 2 33 Index Terms - Machine Learning, Fingerprints, Neural Networks, Deep Learning 34 35

36 The proliferation of large-scale biometric systems has established them as foundational 37 components of modern civil identification, border control, and national security infrastructure. 38 Multimodal databases, such as those maintained by the U.S. Department of Homeland Security 39 (DHS) or the national Aadhaar system in India, house hundreds of millions of identity records, 40 including face, iris, and fingerprint modalities, with volumes expanding rapidly [1-2]. This 41 immense reliance on biometric data quality, however, is compromised by inherent challenges in 42 data acquisition [3]. 43 The central issue is the pervasive problem of data integrity failure stemming from human 44 operational error during enrollment. High throughput demands, coupled with insufficient 45 operator training, frequently result in critical image classification and labeling inaccuracies 46 within these centralized archives. Concrete examples of these errors, documented in law 47 enforcement and government systems (e.g., EBTS records), include: the incorrect sequencing of 48 finger images, intra-modality misclassification (such as labeling a rolled fingerprint as a flat 49 impression), and inter-modality errors (inserting a face or iris image into a fingerprint field) [4-50 6]. The presence of inaccurate data has been shown to degrade the reliability of automated 51 systems, leading to high false rejection rates and the risk of wrongful identification or denial of 52 service [4]. Consequently, ensuring the quality and correct classification of raw data is 53 paramount for maintaining the efficacy of the entire biometric ecosystem. 54 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 3 Currently, the process of rectifying these enrollment errors commonly referred to as de-55 duplication or sequence checking remains a largely manual, tedious, and unscalable task [7-8]. A 56 manual quality assurance process is simply unsustainable for databases containing millions of 57 records. This research addresses this critical gap by developing an automated, deep learning 58 solution for identifying misclassified images, thereby reducing the immense time and cost 59 associated with human inspection [9-17]. 60 The human ability to differentiate and classify fingerprints based on their unique ridge patterns 61 provides the conceptual basis for our automated approach. Expert latent print examiners, through 62 sufficient training, are demonstrably capable of sorting prints based on specific finger types [17]. 63 Similarly, Convolutional Neural Networks (CNNs) are highly suitable for this task because 64 fingerprints possess a specific, repetitive composition of minutiae and ridges [9-10]. Deep 65 learning architectures can effectively learn the statistical characteristics of these patterns, 66 offering a robust alternative to conventional minutiae matching techniques that are sensitive to 67 noise and computational expense [9]. 68 The goal of this research effort is to apply deep learning to automatically detect and flag 69 misclassified biometric images in large-scale datasets for subsequent examination and manual 70 correction. Our primary focus is on intra-modality classification of fingerprints, specifically 71 addressing 20 classes defined by finger identity (thumb, index, etc.) and collection method (flat 72 vs. rolled). We utilize a modified Residual Network (ResNet) architecture in a Siamese 73 configuration to generate robust image embeddings for comparison. Therefore, the specific 74

of this study are to (i) preprocess and organize fingerprint data into 20 distinct classes 75 to create a viable dataset for testing; (ii) design and implement a Siamese network architecture 76 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 4 fine-tuned for high-precision fingerprint classification; (iii) systematically test the network's 77 performance using a 3×4 full factorial design, analyzing the impact of different embedding 78 dimensions (64, 128, 256, 512) and similarity thresholds (0.5, 0.2, 0.1) on classification metrics; 79 and (iv) provide a proof-of-concept tool that quantifies the trade-off between increased 80 conditional classification accuracy and the corresponding rise in the "uncertain" image 81 classification rate, thus guiding practical application for error correction in operational datasets. 82 83 Preliminary Work 84 The foundational concept for the automated biometric data classification tool originated from a 85 preliminary system designed to handle multiple modalities, encompassing face, iris, and 86 fingerprint images. A dedicated preprocessing module was developed to standardize inputs, 87 converting raw WSQ files obtained from data collections into a 180 × 180 × 3 format compatible 88 with the chosen neural network architecture. This initial system utilized a Residual Network 89 (ResNet), pre-trained on the external ImageNet dataset, whose weights were then fine-tuned for 90 biometric sorting over five epochs via transfer learning. Data used for both inter-modality 91 (modality type) and intra-modality (within-modality errors) activities were sourced from 92 numerous collections at the West Virginia University Biometrics Lab under approved IRB 93 protocols. The specific intra-modality tasks included: classifying face images by pose (frontal, 94 profile, other), iris images by side (left or right), and fingerprint images into 20 classes defined 95 by both finger type (index, middle, ring, little, thumb) and collection methodology (flat or rolled 96 impressions). A compilation of the initial results is summarized in Figure 1. 97 98 99 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 5 100 101 Figure 1 Face, Fingerprint and Iris Classification 102 103 The system achieved exceptional performance on high-level tasks: the inter-modality 104 classification of face/fingerprint/iris obtained an accuracy of 99%. Similarly, simple intra-105 modality sorting tasks, such as distinguishing left/right iris images and face pose classification, 106 also yielded an accuracy of 99% (Figure 2). However, the crucial, fine-grained 20-class 107 fingerprint classification task resulted in a markedly lower accuracy of 84. This performance 108 deficit was critically analyzed and attributed to the network’s poor ability to extract sufficiently 109 discriminative features to resolve minute differences between visually similar classes, causing 110 significant confusion between sets like the left/right index and right/left thumbs. As illustrated by 111 the Misclassified Fingerprints from Initial Testing (Figure 3), this high rate of misclassification 112 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 6 for similar prints indicated the need for a network architect ure specialized in metric learning.113 Consequently, the research pivot was made to a different neural network paradigm to improve114 the efficacy of fingerprint detection, specifically aimed at realizing the concept of a classification115 tool (Figure 4) that can automatically flag errors for necessary manual correction. 116 117 118 119 120 121 122 123 124 125 126 127 Figure 2. 128 Iris 129 Classification Results 130 131 132 133 134 135 136 137 138 139 LeftRight 6 g. ve on .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 7 140 141 142 143 Figure 3. Misclassified Fingerprints from Initial Testing. 144 145 146 147

148 149 A. Neural Networks Background 150 The field of Machine Learning (ML) constitutes an intersection of computer science, artificial 151 intelligence, and statistics, fundamentally focused on developing computational systems that 152 automatically improve performance through empirical experience [11]. Neural Networks (NNs) 153 represent a specific, highly effective class of ML models, drawing conceptual inspiration from 154 biological neural systems. These networks operate through interconnected nodes (neurons) that 155 process and transform input data via an adjustable set of weights and biases, guided by 156 algorithms. This concept, rooted in the principle of Hebbian learning that repeated activation 157 strengthens neural connections [12] was first formalized computationally with the Perceptron in 158 the 1950s. Although the field experienced an early recession, innovations in the 1980s, 159 particularly the introduction of backpropagation and gradient descent, coupled with exponential 160 advances in computing power (aligned with Moore’s Law), propelled NNs into a dominant 161 paradigm [13]. In the contemporary era, NNs are essential for complex pattern recognition tasks 162 across all sectors, making them highly pertinent for the specialized image classification required 163 in biometric data analysis. 164 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 8 B. Fingerprint Classification Background 165 Fingerprints represent one of the most reliable modalities in biometric systems, characterized by 166 a unique sequence of ridges and furrows on the finger surface. The core structure is defined by 167 distinct patterns such as the arch (ridges entering one side and exiting the other), loop (ridges 168 curving and exiting the same side they entered), and whorl (ridges forming circular shapes) [14]. 169 These macroscopic patterns are further detailed by minutiae, which are the local irregularities of 170 the ridges, such as the ridge ending (where a ridge terminates) and the ridge bifurcation (where a 171 ridge splits into two) [15]. The uniqueness and consistency of these patterns form the basis for 172 identification, and fingerprint classification based on these feature types is a mature and well-173 understood field [16], with robust feature extractors developed to accurately capture salient 174 characteristics from images [10]. 175 Crucially, this research leverages the concept of human-level expertise in fingerprint analysis. 176 Studies have demonstrated that expert latent print examiners, through sufficient training and 177 experience, develop the ability to accurately distinguish between finger types (e.g., thumb vs. 178 index) [17]. This trained human capability to sort prints based on their unique characteristics 179 beyond core pattern types provides the direct conceptual foundation for our automated approach. 180 By training a neural network to identify and correctly classify prints based on these distinct 181 finger characteristics, we aim to replicate this expert sorting ability to identify subtle 182 classification errors within large datasets [18]. 183 184 C. Image Classification 185 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 9 Image classification is a critical application area for modern machine learning, deeply relevant to 186 this work's objective of classifying biometric images. The standard architecture for analyzing and 187 deciphering visual data is the Convolutional Neural Network (CNN) [19]. Unlike traditional 188 multi-layer feed-forward networks, which struggle with the computational complexity of high-189 resolution image inputs, a CNN effectively manages large pixel densities by leveraging its 190 architectural components to learn hierarchical features [20]. This is achieved by assigning 191 learnable parameters, weights and biases, to automatically extract characteristics that distinguish 192 one image class from another. The core of the CNN is the convolutional layer, which generates a 193 feature map representing specific features extracted across all locations of the input image [19]. 194 This localization and weight-sharing mechanism significantly reduces the number of free 195 parameters compared to fully connected layers, enabling the network to scale to massive 196 datasets. Following the convolutional layer, pooling layers are used to perform dimensionality 197 reduction on the feature maps, optimizing computational efficiency and promoting robustness to 198 minor spatial variations [21-22]. Two common methods are Max Pooling (which retains the 199 maximum value within a defined kernel) and Average Pooling (which computes the average 200 value within the kernel). This hierarchical feature extraction makes CNNs highly effective and 201 computationally feasible for high-accuracy classification tasks, such as those required for 202 fingerprint analysis (Figure 7). 203 204 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 10 205 Figure 4 Max Pooling Demonstration [31] 206 207 With the rapid progress in deep learning, advanced architectures have propelled image208 classification to new levels of performance and applicability [23 ]. For the purposes of this209 research, we focus on two of the most recognized and influential Convolutional Neural Network210 (CNN) architectures: ResNet and the Siamese Network. The Residual Network (ResNet) family,211 particularly ResNet50, is celebrated for its ability to train models with hundreds of layers while212 maintaining high accuracy, primarily through the innovative use of skip connections. This makes213 ResNet a robust base for complex feature extraction in large biometric datasets. This architecture214 provides the necessary foundation for the subsequent implementation of the Siamese network ,215 which is specifically designed for metric learning essential to our classification and error216 detection task. 217 218 D. ResNet50 Architecture 219 ResNet50 is a specific Residual Network architecture introduced by Microsoft Research that220 achieved distinction by winning the ImageNet Large Scale Visual Recognition Challenge221 10 ge his rk ly, ile es re , ror at ge .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 11 (ILSVRC) for its classification efficiency on massive datasets (Figure 8) [21]. The primary 222 innovation of ResNet is its ability to successfully train models with depths ranging from 223 hundreds to thousands of layers while maintaining, and even increasing, accuracy. Historically, 224 merely adding more layers to a standard Deep Convolutional Neural Network (CNN) led to the 225 problem of degradation, where accuracy would saturate and then rapidly decrease due to 226 convergence difficulties and optimization problems [24]. 227 Residual Networks solve this critical challenge through the use of skip connections (or identity 228 mappings). Instead of forcing the network to learn the entire function layer by layer, a skip 229 connection creates an alternate shortcut that bypasses one or more layers, allowing the output 230 from a previous layer to be added directly to the output of a later stacked layer. This formulation 231 ensures that the network is learning the residual mapping rather than the original complex 232 mapping. This "skipping" mechanism facilitates easier gradient flow during backpropagation and 233 enables deeper layers to perform at least as well as their shallower counterparts. This 234 architectural efficiency makes ResNet an ideal feature extraction backbone for applications like 235 our fingerprint classification, where deep learning is used to efficiently sort through massive 236 datasets and label misclassifications. 237 238 239 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 12 240 241 242 243 244 245 246 247 248 249 250 251 252 253 Figure 5. Resnet50 Architecture [24] 254 255 E. Siamese Network Architecture 256 The Siamese network, a class of neural architectures introduced in the 1990s [25], is specifically 257 designed for similarity learning (or metric learning). It consists of two or more identical 258 subnetworks that share the same configuration, parameters, and weights. 259 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 13 260 261 Figure 6. Siamese network architecture 262 263 This structure allows both subnetworks to generate feature vectors (or embeddings) for their 264 respective inputs, which are then compared. Siamese networks are commonly applied to 265 problems such as verification, for instance, determining if two input face images belong to the 266 same individual (Figure 9). In the context of this paper, the same verification principle is applied 267 to the classification of fingerprints [26]. The network is trained using input pairs: a designated 268 anchor image is compared against either a genuine pair (an image belonging to the same finger 269 class) or an impostor pair (an image belonging to a different finger class). The anchor image 270 essentially functions as a representative for its specific class. By comparing the anchor's 271 embedding against all other prints in a dataset, the network generates a quantifiable distance 272 metric that estimates the closeness or distance of other prints to that representative class, 273 enabling fine-grained class separation. This comparative training mechanism is used to 274 determine whether a given print matches the anchor's identity. 275 276 277 278 Shutterstock .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 14 Metrics 279 280 F. Outcome Metrics 281 282 Outcome metrics are a means to quantitatively assess the results of the experiment. In this work, 283 various outcome metrics were used to assess the usability of the neural network. 284 i) Accuracy 285 Accuracy is a metric that allows to evaluate our effectiveness for classification models. It is the 286 number of correct predictions over the total number of predictions [27]. 287 /g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404 /g3015/g3048/g3040/g3029/g3032/g3045 /g3042/g3033 /g3030/g3042/g3045/g3045/g3032/g3030/g3047 /g3043/g3045/g3032/g3031/g3036/g3030/g3047/g3036/g3042/g3041/g3046 /g3021/g3042/g3047/g3028/g3039 /g3041/g3048/g3040/g3029/g3032/g3045 /g3042/g3033 /g3043/g3045/g3032/g3031/g3036/g3030/g3047/g3036/g3042/g3041/g3046 (1) 288 Accuracy is calculated by positive and negative terms. There is True Positive (TP) where the 289 model predicts the positive class correctly. True Negative (TN) is where the model predicts the 290 negative class correctly. False Positive (FP) is when the model predicts a class to be correct 291 when it is not. False Negative (FN) is when the model predicts a class to be incorrect when it is 292 not. By combining these terms, it is possible to obtain a numerical value for accuracy: 293 /g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404 /g3021/g3017/g2878/g3021/g3015 /g3021/g3017/g2878/g3021/g3015/g2878/g3007/g3017/g2878/g3007/g3015 (2) 294 In the case of this research, accuracy defines the number of fingerprints correctly classified over 295 the total number of fingerprints in the dataset. 296 297 2) Conditional and Unconditional Accuracy 298 In the case of this research, there are different ways of calculating the accuracy as it is not simple 299 as the case of only positives and negatives. In order to provide a full picture, we calculated 300 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 15 conditional and unconditional accuracy. Conditional accuracy only considers the images that are 301 classified as positive or negative but does not include the images that are uncertain. 302 Unconditional accuracy on the other hand considers all images/factors. 303 /g1829/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404 /g1846 /g1842/g3397/g1846 /g1840 /g1846/g1842 /g3397 /g1846/g1840 /g3397 /g1832/g1842 /g3397 /g1832/g1840 Equation 1. Conditional Accuracy Equation 304 305 /g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404 /g1846/g1842 /g3397 /g1846/g1840 /g1846/g1842 /g3397 /g1846/g1840 /g3397 /g1832/g1842 /g3397 /g1832/g1840 /g3397 /g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1842/g1867/g1871/g1861/g1872/g1861/g1874/g1857 /g3397 /g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1840/g1857/g1859/g1853/g1872/g1861/g1874/g1857 Equation 2. Unconditional Accuracy Equation 306 307 3) Uncertain 308 Uncertain are predictions that fall in the range of threshold and 1-threshold. These are 309 predictions for which there is not enough sufficient confidence to mark them as positive or 310 negative. These can be images that have been tampered with (blurred or lower quality) or have 311 some sort of error that the neural network is unable to classify accurately. 312 4) Precision 313 Precision determines which proportion of positive identifications were accurately predicted to be 314 correct. It is when True Positive is over the combination of True Positive and False Positive [28]. 315 /g1842/g1870/g1857/g1855/g1861/g1871/g1861/g1867/g1866 /g3404 /g1846/g1842 /g1846 /g1842/g3397/g1832 /g1842 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 16 Equation 3. Precision Equation 316 In the case of this research, precision defines the network’s ability to correctly predict and 317 classify specific classes of the fingerprints, such as front left index and other classes. 318 5) Recall 319 Recall determines which proportion of actual positive identifications were correct. It is 320 calculated using True Positive over the combination of True Positive and False Negative [28]. 321 /g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842 /g1846 /g1842/g3397/g1832 /g1840 Equation 4. Recall Equation 322 323 6) Conditional and Unconditional Recall 324 In the case of this research, there are different ways of calculating the recall as it is not simple as 325 the case of only positives and negatives. In order to provide a full picture, we calculated 326 conditional and unconditional recall. Conditional recall only considers the images that are 327 classified as positive or negative but does not include the images that are uncertain. 328 Unconditional recall on the other hand considers all images. 329 /g1829/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842 /g1846/g1842 /g3397 /g1832/g1840 Equation 5. Conditional Recall 330 331 332 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 17 /g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842 /g1846/g1842 /g3397 /g1832/g1840 /g3397 /g1847/g1866/g1855/g1867/g1866/g1856/g1872/g1861/g1867/g1866/g1853/g1864 /g1842/g1867/g1871/g1861/g1872/g1861/g1874/g1857 Equation 6. Unconditional Recall 333 334 7) F1 Score 335 The F1 Score, also known as the F-measure, is a metric which is based on error. It measures the 336 neural network model’s performance by calculating the harmonic mean of precision and recall 337 for the minority positive class [29]. It is one of the most commonly used metrics for 338 classification models as it provides easy to understand results for balanced and imbalanced 339 datasets factoring in the precision and recall values. 340 /g18321 /g1845/g1855/g1867/g1870/g1857 /g3404 2 /g3400 /g1842/g1870/g1857/g1855/g1861/g1871/g1861/g1867/g1866 /g3400 /g1844/g1857/g1855/g1853/g1864/g1864 /g1842/g1870/g1857/g1855/g1861/g1871/g1861/g1867/g1866 /g3397 /g1844/g1857/g1855/g1853/g1864/g1864 Equation 7. F1 Score Formula 341 To interpret the score, F1 provides an overall model performance from 0 to 1, with 1 being the 342 best possible score. It shows the model’s ability to detect positive cases in recall and accurately 343 classified cases in precision. In the scope of this research, there will be Conditional F1 and 344 Unconditional F1, with one considering the uncertain factor while the other does not (Table 1). 345 Table 1. F1 Score Distribution 346 F1 Score Interpretation of Score Greater than 0.9 Excellent 0.8 – 0.9 Great 0.5 – 0.8 Average/Medium Less than 0.5 Low/Undesirable .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 18 Experiment 347 348 Initial Baseline Classification using ResNet50 349 350 For the biometric data classification effort described here, all raw fingerprint files in the WSQ format 351 were preprocessed and uniformly converted into an 8-bit, 128 × 128 × 3 PNG image structure. The 352 initial classification was centered on a Residual Network (ResNet50) architecture. Network weights 353 were initialized via transfer learning from a model pre-trained on the ImageNet Large Scale Visual 354 Recognition Challenge (ILSVRC) dataset, which encompasses the classification of 1,000 diverse 355 object categories [3]. The transfer learning process involved fine-tuning these weights for the specific 356 biometric sorting purposes, utilizing 5 training epochs. 357 The data utilized was divided into two distinct sets (i) Inter-modality classification was trained and 358 tested on a dedicated, small subset of operational data, comprising 7,500 images (2,500 for each 359 modality: face, iris, and fingerprint). (ii) Intra-modality activities (the primary focus) used a large 360 proprietary dataset collected by the West Virginia University Biometrics Lab under approved IRB 361 protocols, consisting of 40,353 images for training and 10,088 images for testing. The fingerprint 362 images for the intra-modality task were uniformly processed to grayscale and organized into 20 363 distinct classes. These classes are defined by the cross-product of the 10 unique finger identities 364 (left/right thumb, index, middle, ring, little) and the two collection methodologies (flat or rolled 365 impressions). 366 367 ResNet50 Architecture and Baseline Performance 368 The initial 128 × 128 × 3 image input was passed through the ResNet50 network. The network's core 369 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 19 feature extraction was accomplished by the Global Average Pooling layer, which computes the 370 average feature map value across the spatial dimensions. This extraction pipeline then fed into a 371 dense layer of 256 neurons, which served as the image embedding (a reduced-dimensionality vector 372 representation). The final layer was a classification head consisting of 20 neurons, each 373 corresponding to one of the 20 possible finger classes. The network’s output used a one-hot encoding 374 scheme, setting only one output neuron to 1 and the rest to 0 to indicate the predicted class (e.g., a 375 "flat left index" image would correspond to its specific class neuron). The overall classification 376 accuracy for this initial ResNet50 network was 87%. While adequate for simple image classification, 377 this accuracy level was deemed insufficient for reliably scouring large operational datasets where a 378 high degree of certainty is required for flagging misclassified records, thus necessitating an improved 379 network architecture. 380 381 382 383 384 385 386 Figure 7: Siamese Network Architecture 387 388 To achieve more robust and accurate classification performance, the base ResNet50 feature 389 extractor was adapted into a Siamese network configuration, as conceptually illustrated in Figure 390 10. 391 This architecture consists of two identical subnetworks, each acting as a weight-sharing "finger 392 sequence network." Each subnetwork is modified to produce a 256-dimensional embedding 393 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 20 vector for its respective input image. The primary function of these embeddings is to serve as 394 low-dimensional vector representations that encode the salient features of the image, thus 395 reducing data dimensionality while preserving discriminative information. Following the feature 396 extraction, the two embedding vectors are passed to the Euclidean Distance layer. This metric 397 serves as a distance measure, quantifying the relative dissimilarity between the two inputs by 398 calculating the L2 norm between the two 256-dimensional vectors. This distance value is then 399 fed into a sigmoid function. The sigmoid output produces a probability of matching ranging 400 between 0 and 1, where a value near 0 indicates high similarity (the images belong to the same 401 class) and a value near 1 indicates low similarity (the images belong to different classes or 402 classes). 403 404 405 406 Figure 8. An initial architecture 407 408 The initial network architecture relied on a 20-dimensional output vector where each element 409 was mapped to a specific classification class. In contrast, the Siamese network adopts a metric 410 learning paradigm and is trained using input pairs, as depicted in Figure 11. A representative 411 image, referred to as the anchor, is fed into one side of the network. The network's training 412 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 21

is to learn an embedding space where the genuine pair (an image of the same class as 413 the anchor) yields a distance close to 0, and the impostor pair (an image that is not of the same 414 class as the anchor) yields a distance close to 1. Specifically, the network is trained with positive 415 pairs, where the second image belongs to the same class as the anchor, and negative pairs, where 416 the second image belongs to a different class (Figure 12). This comparative training mechanism 417 forces the network to create highly discriminative embeddings that maximize the separation 418 between inter-class feature vectors while minimizing the distance between intra-class feature 419 vectors. 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 Figure 9. Example of Positive Pair (Index and Index) with Negative Pair (Index and Thumb) 436 437 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 22 438 In this study, a 3 × 4 full factorial experimental design was implemented to rigorously evaluate the 439 sensitivity of the Siamese network's performance to two key hyper-parameters: the embedding 440 dimension and the similarity threshold. This systematic approach allowed us to quantify the impact 441 of these components on the network's precision, conditional accuracy, unconditional accuracy, 442 conditional recall, unconditional recall, and the resulting uncertainty factor. Specifically, we tested 443 four embedding dimensions (64, 128, 256, and 512) against three similarity thresholds (0.1, 0.2, and 444 0.5). The inclusion of both the conditional and unconditional variants of accuracy and recall was 445 crucial to precisely determine the magnitude of the "uncertain factor", the percentage of images that 446 fall within the threshold boundary and require manual intervention. To ensure statistical rigor and 447 mitigate the effect of random initialization, each experimental condition was replicated 10 times. The 448 resulting mean values and standard deviations are comprehensively summarized in Figure 13 and 449 Figure 14. 450 451 (a) (b) (c) .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 23 452 Figure 10: Impact of Embeddings on Different Metrics: (a) Conditional Accuracy, (b) 453 Unconditional Accuracy, (c) Precision, (d) Unconditional Recall, and (e) Conditional Recall. 454 455 456 Figure 11: Impact of Thresholds on Different Metrics: (a) Conditional Accuracy, (b) Precision, 457 (c) Unconditional Recall, (d) Unconditional Accuracy, (e) Conditional Recall. 458 459 460 The analysis of the experimental results demonstrates the differential impact of the tested hyper-461 parameters. Varying the embedding dimensions (64, 128, 256, and 512) had no statistically 462 significant impact on classification accuracy or the other evaluated metrics, as evidenced across 463 Figure 2 (a-e). This suggests that a 64-dimensional embedding is sufficient for feature 464 discrimination within this specific metric space. In sharp contrast, modifying the similarity 465 threshold (using subsets of 0.1, 0.2, and 0.5) yielded a significant and systematic change in 466 performance. Narrowing the threshold (i.e., decreasing the value from 0.5 to 0.1) consistently 467 resulted in a substantial increase in conditional accuracy and recall (Figure 14 (a-e)). Crucially, 468 (d) (e) (a) (b) (c) (d) (e) .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 24 this improvement simultaneously increased the uncertain factor from 0% (at a 0.5 threshold) to 469 24% (at a 0.1 threshold). The uncertain factor explicitly represents the number of image 470 comparisons where the Euclidean distance falls within the range of [Threshold, 1 - Threshold]. 471 These are predictions for which the network lacks sufficient confidence to assign a definitive 472 positive or negative classification, consequently requiring manual review. The impact of this 473 factor is clearly visible when comparing the conditional versus unconditional metrics of accuracy 474 and recall (Figure 13 (a) and (e)), where the introduction of the uncertain images drastically 475 depresses the unconditional values. This systematic dependency means that while smaller 476 threshold values yield higher certainty in the classifications made, they dramatically increase the 477 volume of images labeled as "uncertain" (Figure 15). This creates a direct operational trade-off: 478 achieving a high conditional classification accuracy (e.g., 98%) demands accepting an elevated 479 uncertain rate (e.g., 24%) (Table 2). Enhancement of fingerprint image using multiple filters has 480 also been demonstrated [30]. For practical application in databases containing millions of 481 records, a 24% uncertainty level translates to an immense number of prints that must be 482 manually inspected or passed through more sophisticated, yet time-consuming, secondary 483 classification algorithms. Therefore, successful deployment of this proof-of-concept network 484 requires careful optimization to balance the requirement for high classification rigor against the 485 operational feasibility of managing the resulting uncertainty volume. The combination of a high-486 quality benchmark database and an innovative assessment approach are considered to be a strong 487 foundation for future research, contributing to the development of more reliable and secure 488 biometric systems [31]. Fingerprint matching for noisy and distorted patterns using a Siamese 489 Network with ResNet50 and multihead attention has been reported [32]. FootprintNet, a Siamese 490 network that utilizes pre-trained convolutional neural networks, specifically EfficientNet, 491 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 25 MobileNet, and ShuffleNet, to improve the robustness and accuracy of footprint recognition has 492 recently been proposed [33] 493 494 495 Figure 12. Combined Histogram of Predicted Probabilities of Results. 496 497 498 Table 2. Summary of results of threshold, accuracy, uncertain and recall relationship 499 Threshold Accuracy Uncertain Recall/True Positive Rate 0.5 92.5% 0% 92.9% .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 26 0.2 96.8% 14% 96.4% 0.1 98.4% 24% 98.0% 500 501

502 Fingerprints are indispensable for identification in modern biometrics, yet the integrity of large-503 scale databases is continuously threatened by erroneous classification, a critical concern where a 504 single mislabeled image can compromise an entire archive. To address this, we successfully 505 developed a Siamese network for the precise classification of fingerprints by finger type and 506 collection methodology (flat versus rolled impressions). A central objective was to analyze the 507 practical utility of this tool by systematically testing the effect of various embedding dimensions 508 and similarity thresholds (specifically 0.5, 0.2, and 0.1) on performance metrics. Our primary 509 finding reveals a crucial trade-off for operational deployment: using a lower similarity threshold 510 significantly boosts conditional classification accuracy (achieving up to 98%) and precision, but 511 this comes at the expense of classifying a much larger percentage of images as "uncertain" (up to 512 24). This forced manual review of millions of uncertain records in large datasets highlights the 513 challenge of balancing optimal statistical accuracy with practical operational efficiency and cost. 514 This research delivers a robust proof-of-concept tool that moves beyond pure statistical accuracy 515 to provide a quantifiable, practical method for identifying and flagging data integrity errors. The 516 framework is modality-agnostic, making it suitable for deployment across face, iris, and other 517 biometric data. Ultimately, this work contributes to establishing a standardized preprocessing 518 procedure, a lightweight and efficient system that can be applied before any large biometric 519 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 27 dataset is finalized, ensuring the accuracy and legitimacy of identity records while concurrently 520 achieving substantial savings in time and operational costs. 521 522 523 Author Contributions 524 JMD and TIS conceived idea; TIS, design, conducted experiments, interpreted data and prepared 525 manuscript; JMD and NMN, supervised, reviewed and improved the manuscript; All authors 526 agreed to submit the manuscript. 527 528

529 TIS is thankful to the funding GRA provided by the JMD during this study. 530 531 Conflict of Interest Statement 532 The authors declare no conflict on interests. 533 Data Availability Statement 534 All data are included in this manuscript. 535 ORCID 536 Tahsin Islam Sakif: 0000-0001-5378-4775 537 N. M. Nasrabadi: 0000-0001-8730-627X 538 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 28 J M Dawson: 0000-0002-4539-7588 539 540 541 542 543

544 [1] J. Rohrlich, “Homeland Security will soon have biometric data on nearly 260 million 545 people,” Quartz, Nov. 07, 2019. https://qz.com/1744400/dhs-expected-to-have-biometrics-on-546 260-million-people-by-2022 (accessed Mar. 21, 2023). 547 [2] U. Rao and V. Nair, “Aadhaar: Governing with Biometrics,” South Asia: Journal of South 548 Asian Studies , vol. 42, no. 3, pp. 469–481, May 2019, doi: 549 https://doi.org/10.1080/00856401.2019.1595343. 550 [3] J. Lynch, “HART: Homeland Security’s Massive New Database Will Include Face 551 Recognition, DNA, and Peoples’ ‘Non-Obvious Relationships,’” Electronic Frontier 552 Foundation, Jun. 07, 2018. https://www.eff.org/deeplinks/2018/06/hart-homeland-securitys-553 massive-new-database-will-include-face-recognition-dna-and 554 [4] H. Rudolph, L. Moy, and A. Bedoya, “Airport Face Scans: An Investigation,” 555 www.airportfacescans.com, Dec. 17, 2017. https://www.airportfacescans.com/ (accessed Mar. 556 21, 2023). 557 [5] P. Higgins, “Standards for the electronic submission of fingerprint cards to the FBI,” Journal 558 of Forenseic Identification, vol. 45, no. 4, pp. 409–418, 1995. 559 [6] K. E. Wertheim, “Human Factors in Large-Scale Biometric Systems: A Study of the Human 560 Factors Related to Errors in Semiautomatic Fingerprint Biometrics,” IEEE Systems Journal, vol. 561 4, no. 2, pp. 138–146, Jun. 2010, doi: https://doi.org/10.1109/jsyst.2010.2049878. 562 [7] “What Is Data Deduplication? | Benefits & Use Cases | NetApp,” www.netapp.com. 563 https://www.netapp.com/data-management/what-is-data-deduplication/ 564 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 29 [8] P. S. Sudhish, A. K. Jain, and K. Cao, “Adaptive fusion of biometric and biographic 565 information for identity de-duplication,” Pattern Recognition Letters, vol. 84, pp. 199–207, Dec. 566 2016, doi: https://doi.org/10.1016/j.patrec.2016.10.011. 567 [9] M. I. Jordan and T. M. Mitchell, “Machine learning: Trends, perspectives, and prospects,” 568 Science, vol. 349, no. 6245, pp. 255–260, Jul. 2015, doi: 569 https://doi.org/10.1126/science.aaa8415. 570 [10] H. Lin and A. Jain, “Classification of fingerprint images,” vol. 2, pp. 665–672, 1999, 571 Available: https://www.researchgate.net/profile/Lin-Hong-572 12/publication/2929466_Classification_of_Fingerprint_Images/links/54733a5a0cf216f8cfaeb8b7573 /Classification-of-Fingerprint-Images.pdf 574 [11] K. Fukushima, “A neural network for visual pattern recognition,” Computer, vol. 21, no. 3, 575 pp. 65–75, Mar. 1988, doi: https://doi.org/10.1109/2.32. 576 [12] D. O. Hebb, The organization of behavior : a neuropsychological theory . New York: 577 Routledge, Taylor & Francis Group, 2012. 578 [13] F. Rosenblatt, The Perceptron . Ithaca, New York: Cornell Aeronautical Laboratory, 1958, 579 pp. 85–4601. 580 [14] Q. Zhang and H. Yan, “Fingerprint classification based on extraction and analysis of 581 singularities and pseudo ridges,” Pattern Recognition , vol. 37, no. 11, pp. 2233–2243, Nov. 582 2004, doi: https://doi.org/10.1016/j.patcog.2003.12.020. 583 [15] D. Michelsanti, A.-D. Ene, Y. Guichi, R. Stef, K. Nasrollahi, and T. B. Moeslund, “Fast 584 Fingerprint Classification with Deep Neural Networks,” Proceedings of the 12th International 585 Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and 586 Applications, 2017, doi: https://doi.org/10.5220/0006116502020209. 587 [16] B. Pandya, G. Cosma, A. A. Alani, A. Taherkhani, V. Bharadi, and T. M. McGinnity, 588 “Fingerprint classification using a deep convolutional neural network,” 2018 4th International 589 Conference on Information Management (ICIM) , pp. 86–91, May 2018, doi: 590 https://doi.org/10.1109/infoman.2018.8392815. 591 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 30 [17] R. A. Searston and J. M. Tangen, “The Emergence of Perceptual Expertise with Fingerprints 592 Over Time,” Journal of Applied Research in Memory and Cognition , vol. 6, no. 4, pp. 442–451, 593 Dec. 2017, doi: https://doi.org/10.1016/j.jarmac.2017.08.006. 594 [18] N. K. Ratha, K. Karu, Shaoyun Chen and A. K. Jain, "A real-time matching system for large 595 fingerprint databases," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 596 18, no. 8, pp. 799-813, Aug. 1996, doi: 10.1109/34.531800. 597 [19] Y. LeCun, K. Kavukcuoglu, and C. Farabet, “Convolutional networks and applications in 598 vision,” Proceedings of 2010 IEEE International Symposium on Circuits and Systems , pp. 253–599 256, May 2010, doi: https://doi.org/10.1109/iscas.2010.5537907. 600 [20] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale 601 Image Recognition,” arXiv.org, 2014. https://arxiv.org/abs/1409.1556 602 [21] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” 603 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, Jun. 604 2016, doi: https://doi.org/10.1109/cvpr.2016.90. 605 [22] V. Meshram and K. Patil, “Border-Square net: a robust multi-grade fruit classification in 606 IoT smart agriculture using feature extraction based Deep Maxout network,” Multimedia Tools 607 and Applications, May 2022, doi: https://doi.org/10.1007/s11042-022-12855-7. 608 [23] Q. Zheng, M. Zhang, and J. Yang, “A Bilinear Multi-Scale Convolutional Neural Network 609 for Fine-grained Object Classification,” IAENG International Journal of Computer Science, vol. 610 45, no. 2, 2018, Available: https://www.iaeng.org/IJCS/issues_v45/issue_2/IJCS_45_2_12.pdf 611 [24] K. Sai Prasad and D. Pasupathy, “DEEP CONVOLUTIONAL NEURAL NETWORK 612 IMPLEMENTATIONS FOR EFFICIENT BRAIN STROKE DETECTION USING MRI 613 SCANS,” Journal of Theoretical and Applied Sciences , vol. 100, no. 17, 2022, Accessed: Apr. 614 24, 2023. [Online]. Available: http://www.jatit.org/volumes/Vol100No17/14Vol100No17.pdf 615 [25] J. Bromley et al. , “SIGNATURE VERIFICATION USING A ‘SIAMESE’ TIME DELAY 616 NEURAL NETWORK,” International Journal of Pattern Recognition and Artificial 617 Intelligence, vol. 07, no. 04, pp. 669–688, Aug. 1993, doi: 618 https://doi.org/10.1142/s0218001493000339. 619 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint 31 [ 2 6 ] S . K . R o y , M . H a r a n d i , R . N o c k , a n d R . H a r t l e y , “ S i a m e s e N e tw o r k s : T h e T a l e o f T w o 620 Manifolds,” openaccess.thecvf.com, 2019. 621 https://openaccess.thecvf.com/content_ICCV_2019/html/Roy_Siamese_Networks_The_Tale_of622 _Two_Manifolds_ICCV_2019_paper.html (accessed Apr. 24, 2023). 623 [27] Google, “Classification: Accuracy | Machine Learning Crash Course,” Google Developers, 624 2019. https://developers.google.com/machine-learning/crash-course/classification/accuracy 625 [28] Google, “Classification: Precision and Recall | Machine Learning Crash Course 626 | Google Developers,” Google Developers , Mar. 05, 2019. 627 https://developers.google.com/machine-learning/crash-course/classification/precision-and-recall 628 [29] J. Korstanje, “The F1 score,” Medium, Aug. 31, 2021. https://towardsdatascience.com/the-629 f1-score-bec2bbc38aa6 630 [30] Shams H, Jan T, Khalil AA, Ahmad N, Munir A, Khalil RA. 2023. Fingerprint image 631 enhancement using multiple filters. PeerJ Computer Science 9:e1183 632 https://doi.org/10.7717/peerj-cs.1183 633 [31] Bouhamed, S.A. Exploring possibilistic fingerprint image quality analysis: a soft computing 634 approach for biometric database management. Int. J. Mach. Learn. & Cyber. (2025). 635 https://doi.org/10.1007/s13042-025-02587-z 636 [32] T. Sawhney et al., Fingerprint matching for noisy and distorted patterns using a Siamese 637 Network with ResNet50 and multihead attenti on. IEEE Access, vol. 13, pp. 88047-88064, 2025, 638 doi: 10.1109/ACCESS.2025.3567328. 639 [33] İ brahimoğ lu, N., Osmani, A., Ghaffari, A. et al. FootprintNet: a Siamese network method for 640 biometric identification using footprints. J Supercomput 81, 714 (2025). 641 https://doi.org/10.1007/s11227-025-07170-5 642 643 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint