{"paper_id":"0ee82c7b-64e1-482f-bce9-e7db37b24a5b","body_text":"1 \n \nFinger Type Classification for Fingerprint Image Error Correction 1 \nin Large Scale Biometric Databases 2 \n 3 \nTahsin Islam. Sakif, Nasser Nasrabadi and Jeremy Dawson* 4 \nWest Virginia University 5 \nStatler College of Engineering and Mineral Resources 6 \nLane Department of Computer Science and Electrical Engineering 7 \nP.O. Box 6109, Morgantown, WV 26506 8 \n*Corresponding author: jeremy.dawson@mail.wvu.edu 9 \n 10 \nABSTRACT 11 \n 12 \nAbstract - Large-scale biometric systems, essential for national security and border 13 \nmanagement, increasingly rely on multimodal databases containing millions of identities. 14 \nHowever, operational pressures and insufficient training lead to frequent image classification and 15 \nlabeling errors by human operators. These critical data integrity issues include the mislabeling of 16 \nrolled vs. flat fingerprints, out-of-sequence captures, and the insertion of incorrect modalities. 17 \nSuch errors render enrollment records unreliable, compromising subsequent identity verification 18 \nprocesses. Since manually sorting vast image archives is unfeasible, our study proposes an 19 \nautomated solution. The primary objective was to deploy a Siamese Network to classify 20 \nfingerprints by their precise finger type and collection methodology (flat or rolled impressions). 21 \nA secondary, but central, goal was to investigate the influence of varying embedding dimensions 22 \n(64, 128, 256, 512) and similarity thresholds (0.5, 0.2, 0.1) on the network's performance 23 \nmetrics. Our most significant finding demonstrates a clear trade-off: a lower similarity threshold 24 \ndrastically increases conditional accuracy and precision (e.g., up to 98%) but simultaneously 25 \nincreases the proportion of images categorized as \"uncertain\" (up to 24%). In a practical, large-26 \nscale application, this necessitates balancing superior classification accuracy against a higher 27 \nvolume of images requiring costly manual inspection. This work provides a proof-of-concept 28 \ntool capable of efficiently quantifying the percentage of images requiring human review across 29 \nvarious modalities (fingerprints, face, iris). The eventual goal is a lightweight, efficient tool to 30 \nestablish standard preprocessing procedures for any large biometric dataset, dramatically 31 \nreducing the time and cost associated with data integrity maintenance. 32 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n2 \n \n 33 \nIndex Terms - Machine Learning, Fingerprints, Neural Networks, Deep Learning 34 \n 35 \nIntroduction 36 \nThe proliferation of large-scale biometric systems has established them as foundational 37 \ncomponents of modern civil identification, border control, and national security infrastructure. 38 \nMultimodal databases, such as those maintained by the U.S. Department of Homeland Security 39 \n(DHS) or the national Aadhaar system in India, house hundreds of millions of identity records, 40 \nincluding face, iris, and fingerprint modalities, with volumes expanding rapidly [1-2]. This 41 \nimmense reliance on biometric data quality, however, is compromised by inherent challenges in 42 \ndata acquisition [3]. 43 \nThe central issue is the pervasive problem of data integrity failure stemming from human 44 \noperational error during enrollment. High throughput demands, coupled with insufficient 45 \noperator training, frequently result in critical image classification and labeling inaccuracies 46 \nwithin these centralized archives. Concrete examples of these errors, documented in law 47 \nenforcement and government systems (e.g., EBTS records), include: the incorrect sequencing of 48 \nfinger images, intra-modality misclassification (such as labeling a rolled fingerprint as a flat 49 \nimpression), and inter-modality errors (inserting a face or iris image into a fingerprint field) [4-50 \n6]. The presence of inaccurate data has been shown to degrade the reliability of automated 51 \nsystems, leading to high false rejection rates and the risk of wrongful identification or denial of 52 \nservice [4]. Consequently, ensuring the quality and correct classification of raw data is 53 \nparamount for maintaining the efficacy of the entire biometric ecosystem. 54 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n3 \n \nCurrently, the process of rectifying these enrollment errors commonly referred to as de-55 \nduplication or sequence checking remains a largely manual, tedious, and unscalable task [7-8]. A 56 \nmanual quality assurance process is simply unsustainable for databases containing millions of 57 \nrecords. This research addresses this critical gap by developing an automated, deep learning 58 \nsolution for identifying misclassified images, thereby reducing the immense time and cost 59 \nassociated with human inspection [9-17]. 60 \nThe human ability to differentiate and classify fingerprints based on their unique ridge patterns 61 \nprovides the conceptual basis for our automated approach. Expert latent print examiners, through 62 \nsufficient training, are demonstrably capable of sorting prints based on specific finger types [17]. 63 \nSimilarly, Convolutional Neural Networks (CNNs) are highly suitable for this task because 64 \nfingerprints possess a specific, repetitive composition of minutiae and ridges [9-10]. Deep 65 \nlearning architectures can effectively learn the statistical characteristics of these patterns, 66 \noffering a robust alternative to conventional minutiae matching techniques that are sensitive to 67 \nnoise and computational expense [9]. 68 \nThe goal of this research effort is to apply deep learning to automatically detect and flag 69 \nmisclassified biometric images in large-scale datasets for subsequent examination and manual 70 \ncorrection. Our primary focus is on intra-modality classification of fingerprints, specifically 71 \naddressing 20 classes defined by finger identity (thumb, index, etc.) and collection method (flat 72 \nvs. rolled). We utilize a modified Residual Network (ResNet) architecture in a Siamese 73 \nconfiguration to generate robust image embeddings for comparison. Therefore, the specific 74 \nobjectives of this study are to (i) preprocess and organize fingerprint data into 20 distinct classes 75 \nto create a viable dataset for testing; (ii) design and implement a Siamese network architecture 76 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n4 \n \nfine-tuned for high-precision fingerprint classification; (iii) systematically test the network's 77 \nperformance using a 3×4 full factorial design, analyzing the impact of different embedding 78 \ndimensions (64, 128, 256, 512) and similarity thresholds (0.5, 0.2, 0.1) on classification metrics; 79 \nand (iv) provide a proof-of-concept tool that quantifies the trade-off between increased 80 \nconditional classification accuracy and the corresponding rise in the \"uncertain\" image 81 \nclassification rate, thus guiding practical application for error correction in operational datasets. 82 \n 83 \nPreliminary Work 84 \nThe foundational concept for the automated biometric data classification tool originated from a 85 \npreliminary system designed to handle multiple modalities, encompassing face, iris, and 86 \nfingerprint images. A dedicated preprocessing module was developed to standardize inputs, 87 \nconverting raw WSQ files obtained from data collections into a 180 × 180 × 3 format compatible 88 \nwith the chosen neural network architecture. This initial system utilized a Residual Network 89 \n(ResNet), pre-trained on the external ImageNet dataset, whose weights were then fine-tuned for 90 \nbiometric sorting over five epochs via transfer learning. Data used for both inter-modality 91 \n(modality type) and intra-modality (within-modality errors) activities were sourced from 92 \nnumerous collections at the West Virginia University Biometrics Lab under approved IRB 93 \nprotocols. The specific intra-modality tasks included: classifying face images by pose (frontal, 94 \nprofile, other), iris images by side (left or right), and fingerprint images into 20 classes defined 95 \nby both finger type (index, middle, ring, little, thumb) and collection methodology (flat or rolled 96 \nimpressions). A compilation of the initial results is summarized in Figure 1.  97 \n 98 \n 99 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n5 \n \n 100 \n  101 \nFigure 1 Face, Fingerprint and Iris Classification 102 \n 103 \nThe system achieved exceptional performance on high-level tasks: the inter-modality 104 \nclassification of face/fingerprint/iris obtained an accuracy of 99%. Similarly, simple intra-105 \nmodality sorting tasks, such as distinguishing left/right iris images and face pose classification, 106 \nalso yielded an accuracy of 99% (Figure 2). However, the crucial, fine-grained 20-class 107 \nfingerprint classification task resulted in a markedly lower accuracy of 84. This performance 108 \ndeficit was critically analyzed and attributed to the network’s poor ability to extract sufficiently 109 \ndiscriminative features to resolve minute differences between visually similar classes, causing 110 \nsignificant confusion between sets like the left/right index and right/left thumbs. As illustrated by 111 \nthe Misclassified Fingerprints from Initial Testing (Figure 3), this high rate of misclassification 112 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n6\n \nfor similar prints indicated the need for a network architect ure specialized in metric learning.113 \nConsequently, the research pivot was made to a different neural network paradigm to improve114 \nthe efficacy of fingerprint detection, specifically aimed at realizing the concept of a classification115 \ntool (Figure 4) that can automatically flag errors for necessary manual correction. 116 \n 117 \n 118 \n 119 \n 120 \n 121 \n 122 \n 123 \n 124 \n 125 \n 126 \n 127 \nFigure 2. 128 \nIris 129 \nClassification Results 130 \n 131 \n 132 \n 133 \n 134 \n 135 \n 136 \n 137 \n 138 \n 139 \nLeftRight\n6 \ng. \nve \non \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n7 \n \n 140 \n 141 \n 142 \n 143 \nFigure 3. Misclassified Fingerprints from Initial Testing. 144 \n 145 \n 146 \n 147 \nMethods 148 \n 149 \nA. Neural Networks Background 150 \nThe field of Machine Learning (ML) constitutes an intersection of computer science, artificial 151 \nintelligence, and statistics, fundamentally focused on developing computational systems that 152 \nautomatically improve performance through empirical experience [11]. Neural Networks (NNs) 153 \nrepresent a specific, highly effective class of ML models, drawing conceptual inspiration from 154 \nbiological neural systems. These networks operate through interconnected nodes (neurons) that 155 \nprocess and transform input data via an adjustable set of weights and biases, guided by 156 \nalgorithms. This concept, rooted in the principle of Hebbian learning that repeated activation 157 \nstrengthens neural connections [12] was first formalized computationally with the Perceptron in 158 \nthe 1950s. Although the field experienced an early recession, innovations in the 1980s, 159 \nparticularly the introduction of backpropagation and gradient descent, coupled with exponential 160 \nadvances in computing power (aligned with Moore’s Law), propelled NNs into a dominant 161 \nparadigm [13]. In the contemporary era, NNs are essential for complex pattern recognition tasks 162 \nacross all sectors, making them highly pertinent for the specialized image classification required 163 \nin biometric data analysis. 164 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n8 \n \nB. Fingerprint Classification Background 165 \nFingerprints represent one of the most reliable modalities in biometric systems, characterized by 166 \na unique sequence of ridges and furrows on the finger surface. The core structure is defined by 167 \ndistinct patterns such as the arch (ridges entering one side and exiting the other), loop (ridges 168 \ncurving and exiting the same side they entered), and whorl (ridges forming circular shapes) [14]. 169 \nThese macroscopic patterns are further detailed by minutiae, which are the local irregularities of 170 \nthe ridges, such as the ridge ending (where a ridge terminates) and the ridge bifurcation (where a 171 \nridge splits into two) [15]. The uniqueness and consistency of these patterns form the basis for 172 \nidentification, and fingerprint classification based on these feature types is a mature and well-173 \nunderstood field [16], with robust feature extractors developed to accurately capture salient 174 \ncharacteristics from images [10]. 175 \nCrucially, this research leverages the concept of human-level expertise in fingerprint analysis. 176 \nStudies have demonstrated that expert latent print examiners, through sufficient training and 177 \nexperience, develop the ability to accurately distinguish between finger types (e.g., thumb vs. 178 \nindex) [17]. This trained human capability to sort prints based on their unique characteristics 179 \nbeyond core pattern types provides the direct conceptual foundation for our automated approach. 180 \nBy training a neural network to identify and correctly classify prints based on these distinct 181 \nfinger characteristics, we aim to replicate this expert sorting ability to identify subtle 182 \nclassification errors within large datasets [18]. 183 \n 184 \nC. Image Classification 185 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n9 \n \nImage classification is a critical application area for modern machine learning, deeply relevant to 186 \nthis work's objective of classifying biometric images. The standard architecture for analyzing and 187 \ndeciphering visual data is the Convolutional Neural Network (CNN) [19]. Unlike traditional 188 \nmulti-layer feed-forward networks, which struggle with the computational complexity of high-189 \nresolution image inputs, a CNN effectively manages large pixel densities by leveraging its 190 \narchitectural components to learn hierarchical features [20]. This is achieved by assigning 191 \nlearnable parameters, weights and biases, to automatically extract characteristics that distinguish 192 \none image class from another. The core of the CNN is the convolutional layer, which generates a 193 \nfeature map representing specific features extracted across all locations of the input image [19]. 194 \nThis localization and weight-sharing mechanism significantly reduces the number of free 195 \nparameters compared to fully connected layers, enabling the network to scale to massive 196 \ndatasets. Following the convolutional layer, pooling layers are used to perform dimensionality 197 \nreduction on the feature maps, optimizing computational efficiency and promoting robustness to 198 \nminor spatial variations [21-22]. Two common methods are Max Pooling (which retains the 199 \nmaximum value within a defined kernel) and Average Pooling (which computes the average 200 \nvalue within the kernel). This hierarchical feature extraction makes CNNs highly effective and 201 \ncomputationally feasible for high-accuracy classification tasks, such as those required for 202 \nfingerprint analysis (Figure 7). 203 \n 204 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n10\n \n205 \nFigure 4 Max Pooling Demonstration [31] 206 \n 207 \nWith the rapid progress in deep learning, advanced architectures have propelled image208 \nclassification to new levels of performance and applicability [23 ]. For the purposes of this209 \nresearch, we focus on two of the most recognized and influential Convolutional Neural Network210 \n(CNN) architectures: ResNet and the Siamese Network. The Residual Network (ResNet) family,211 \nparticularly ResNet50, is celebrated for its ability to train models with hundreds of layers while212 \nmaintaining high accuracy, primarily through the innovative use of skip connections. This makes213 \nResNet a robust base for complex feature extraction in large biometric datasets. This architecture214 \nprovides the necessary foundation for the subsequent implementation of the Siamese network ,215 \nwhich is specifically designed for metric learning essential to our classification and error216 \ndetection task. 217 \n 218 \nD. ResNet50 Architecture 219 \nResNet50 is a specific Residual Network architecture introduced by Microsoft Research that220 \nachieved distinction by winning the ImageNet Large Scale Visual Recognition Challenge221 \n10 \n \nge \nhis \nrk \nly, \nile \nes \nre \n, \nror \nat \nge \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n11 \n \n(ILSVRC) for its classification efficiency on massive datasets (Figure 8) [21]. The primary 222 \ninnovation of ResNet is its ability to successfully train models with depths ranging from 223 \nhundreds to thousands of layers while maintaining, and even increasing, accuracy. Historically, 224 \nmerely adding more layers to a standard Deep Convolutional Neural Network (CNN) led to the 225 \nproblem of degradation, where accuracy would saturate and then rapidly decrease due to 226 \nconvergence difficulties and optimization problems [24]. 227 \nResidual Networks solve this critical challenge through the use of skip connections (or identity 228 \nmappings). Instead of forcing the network to learn the entire function layer by layer, a skip 229 \nconnection creates an alternate shortcut that bypasses one or more layers, allowing the output 230 \nfrom a previous layer to be added directly to the output of a later stacked layer. This formulation 231 \nensures that the network is learning the residual mapping  rather than the original complex 232 \nmapping. This \"skipping\" mechanism facilitates easier gradient flow during backpropagation and 233 \nenables deeper layers to perform at least as  well as their shallower counterparts. This 234 \narchitectural efficiency makes ResNet an ideal feature extraction backbone for applications like 235 \nour fingerprint classification, where deep learning is used to efficiently sort through massive 236 \ndatasets and label misclassifications. 237 \n  238 \n 239 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n12 \n \n 240 \n 241 \n 242 \n 243 \n 244 \n 245 \n 246 \n 247 \n 248 \n 249 \n 250 \n 251 \n 252 \n 253 \nFigure 5. Resnet50 Architecture [24] 254 \n 255 \nE. Siamese Network Architecture 256 \nThe Siamese network, a class of neural architectures introduced in the 1990s [25], is specifically 257 \ndesigned for similarity learning (or metric learning). It consists of two or more identical 258 \nsubnetworks that share the same configuration, parameters, and weights. 259 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n13 \n \n 260 \n 261 \nFigure 6. Siamese network architecture 262 \n 263 \nThis structure allows both subnetworks to generate feature vectors (or embeddings) for their 264 \nrespective inputs, which are then compared. Siamese networks are commonly applied to 265 \nproblems such as verification, for instance, determining if two input face images belong to the 266 \nsame individual (Figure 9). In the context of this paper, the same verification principle is applied 267 \nto the classification of fingerprints [26]. The network is trained using input pairs: a designated 268 \nanchor image is compared against either a genuine pair (an image belonging to the same finger 269 \nclass) or an impostor pair (an image belonging to a different finger class). The anchor image 270 \nessentially functions as a representative for its specific class. By comparing the anchor's 271 \nembedding against all other prints in a dataset, the network generates a quantifiable distance 272 \nmetric that estimates the closeness or distance of other prints to that representative class, 273 \nenabling fine-grained class separation. This comparative training mechanism is used to 274 \ndetermine whether a given print matches the anchor's identity. 275 \n 276 \n 277 \n 278 \nShutterstock\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n14 \n \nMetrics 279 \n 280 \nF. Outcome Metrics 281 \n 282 \nOutcome metrics are a means to quantitatively assess the results of the experiment. In this work, 283 \nvarious outcome metrics were used to assess the usability of the neural network. 284 \ni) Accuracy 285 \nAccuracy is a metric that allows to evaluate our effectiveness for classification models. It is the 286 \nnumber of correct predictions over the total number of predictions [27]. 287 \n/g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404\n/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\n/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 \nAccuracy is calculated by positive and negative terms. There is True Positive (TP) where the 289 \nmodel predicts the positive class correctly. True Negative (TN) is where the model predicts the 290 \nnegative class correctly. False Positive (FP) is when the model predicts a class to be correct 291 \nwhen it is not. False Negative (FN) is when the model predicts a class to be incorrect when it is 292 \nnot. By combining these terms, it is possible to obtain a numerical value for accuracy: 293 \n/g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877 /g3404\n/g3021/g3017/g2878/g3021/g3015\n/g3021/g3017/g2878/g3021/g3015/g2878/g3007/g3017/g2878/g3007/g3015  (2) 294 \nIn the case of this research, accuracy defines the number of fingerprints correctly classified over 295 \nthe total number of fingerprints in the dataset. 296 \n 297 \n2) Conditional and Unconditional Accuracy 298 \nIn the case of this research, there are different ways of calculating the accuracy as it is not simple 299 \nas the case of only positives and negatives. In order to provide a full picture, we calculated 300 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n15 \n \nconditional and unconditional accuracy. Conditional accuracy only considers the images that are 301 \nclassified as positive or negative but does not include the images that are uncertain. 302 \nUnconditional accuracy on the other hand considers all images/factors.  303 \n/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\n/g1846/g1842 /g3397 /g1846/g1840 /g3397 /g1832/g1842 /g3397 /g1832/g1840  \nEquation 1. Conditional Accuracy Equation 304 \n 305 \n/g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1827/g1855/g1855/g1873/g1870/g1853/g1855/g1877\n/g3404 /g1846/g1842 /g3397 /g1846/g1840\n/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  \nEquation 2. Unconditional Accuracy Equation 306 \n 307 \n3) Uncertain 308 \nUncertain are predictions that fall in the range of threshold and 1-threshold. These are 309 \npredictions for which there is not enough sufficient confidence to mark them as positive or 310 \nnegative. These can be images that have been tampered with (blurred or lower quality) or have 311 \nsome sort of error that the neural network is unable to classify accurately.  312 \n4) Precision 313 \nPrecision determines which proportion of positive identifications were accurately predicted to be 314 \ncorrect. It is when True Positive is over the combination of True Positive and False Positive [28].  315 \n/g1842/g1870/g1857/g1855/g1861/g1871/g1861/g1867/g1866 /g3404 /g1846/g1842\n/g1846 /g1842/g3397/g1832 /g1842  \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n16 \n \nEquation 3. Precision Equation 316 \nIn the case of this research, precision defines the network’s ability to correctly predict and 317 \nclassify specific classes of the fingerprints, such as front left index and other classes. 318 \n5) Recall 319 \nRecall determines which proportion of actual positive identifications were correct. It is 320 \ncalculated using True Positive over the combination of True Positive and False Negative [28].  321 \n/g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842\n/g1846 /g1842/g3397/g1832 /g1840  \nEquation 4. Recall Equation 322 \n 323 \n6) Conditional and Unconditional Recall 324 \nIn the case of this research, there are different ways of calculating the recall as it is not simple as 325 \nthe case of only positives and negatives. In order to provide a full picture, we calculated 326 \nconditional and unconditional recall. Conditional recall only considers the images that are 327 \nclassified as positive or negative but does not include the images that are uncertain. 328 \nUnconditional recall on the other hand considers all images.  329 \n/g1829/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842\n/g1846/g1842 /g3397 /g1832/g1840  \nEquation 5. Conditional Recall 330 \n 331 \n 332 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n17 \n \n/g1847/g1866/g1855/g1867/g1866/g1856/g1861/g1872/g1861/g1867/g1866/g1853/g1864 /g1844/g1857/g1855/g1853/g1864/g1864 /g3404 /g1846/g1842\n/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  \nEquation 6. Unconditional Recall 333 \n 334 \n7) F1 Score 335 \nThe F1 Score, also known as the F-measure, is a metric which is based on error. It measures the 336 \nneural network model’s performance by calculating the harmonic mean of precision and recall 337 \nfor the minority positive class [29]. It is one of the most commonly used metrics for 338 \nclassification models as it provides easy to understand results for balanced and imbalanced 339 \ndatasets factoring in the precision and recall values. 340 \n/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\n/g1842/g1870/g1857/g1855/g1861/g1871/g1861/g1867/g1866 /g3397 /g1844/g1857/g1855/g1853/g1864/g1864  \nEquation 7. F1 Score Formula 341 \nTo interpret the score, F1 provides an overall model performance from 0 to 1, with 1 being the 342 \nbest possible score. It shows the model’s ability to detect positive cases in recall and accurately 343 \nclassified cases in precision. In the scope of this research, there will be Conditional F1 and 344 \nUnconditional F1, with one considering the uncertain factor while the other does not (Table 1). 345 \nTable 1. F1 Score Distribution 346 \nF1 Score Interpretation of Score \nGreater than 0.9 Excellent \n0.8 – 0.9 Great \n0.5 – 0.8  Average/Medium \nLess than 0.5 Low/Undesirable \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n18 \n \nExperiment 347 \n 348 \nInitial Baseline Classification using ResNet50 349 \n 350 \nFor the biometric data classification effort described here, all raw fingerprint files in the WSQ format 351 \nwere preprocessed and uniformly converted into an 8-bit, 128 × 128 × 3 PNG image structure. The 352 \ninitial classification was centered on a Residual Network (ResNet50) architecture. Network weights 353 \nwere initialized via transfer learning from a model pre-trained on the ImageNet Large Scale Visual 354 \nRecognition Challenge (ILSVRC) dataset, which encompasses the classification of 1,000 diverse 355 \nobject categories [3]. The transfer learning process involved fine-tuning these weights for the specific 356 \nbiometric sorting purposes, utilizing 5 training epochs. 357 \nThe data utilized was divided into two distinct sets (i) Inter-modality classification was trained and 358 \ntested on a dedicated, small subset of operational data, comprising 7,500 images (2,500 for each 359 \nmodality: face, iris, and fingerprint). (ii) Intra-modality activities (the primary focus) used a large 360 \nproprietary dataset collected by the West Virginia University Biometrics Lab under approved IRB 361 \nprotocols, consisting of 40,353 images for training and 10,088 images for testing. The fingerprint 362 \nimages for the intra-modality task were uniformly processed to grayscale and organized into 20 363 \ndistinct classes. These classes are defined by the cross-product of the 10 unique finger identities 364 \n(left/right thumb, index, middle, ring, little) and the two collection methodologies (flat or rolled 365 \nimpressions). 366 \n 367 \nResNet50 Architecture and Baseline Performance 368 \nThe initial 128 × 128 × 3 image input was passed through the ResNet50 network. The network's core 369 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n19 \n \nfeature extraction was accomplished by the Global Average Pooling layer, which computes the 370 \naverage feature map value across the spatial dimensions. This extraction pipeline then fed into a 371 \ndense layer of 256 neurons, which served as the image embedding (a reduced-dimensionality vector 372 \nrepresentation). The final layer was a classification head consisting of 20 neurons, each 373 \ncorresponding to one of the 20 possible finger classes. The network’s output used a one-hot encoding 374 \nscheme, setting only one output neuron to 1 and the rest to 0 to indicate the predicted class (e.g., a 375 \n\"flat left index\" image would correspond to its specific class neuron). The overall classification 376 \naccuracy for this initial ResNet50 network was 87%. While adequate for simple image classification, 377 \nthis accuracy level was deemed insufficient for reliably scouring large operational datasets where a 378 \nhigh degree of certainty is required for flagging misclassified records, thus necessitating an improved 379 \nnetwork architecture. 380 \n 381 \n 382 \n 383 \n 384 \n 385 \n 386 \nFigure 7: Siamese Network Architecture 387 \n 388 \nTo achieve more robust and accurate classification performance, the base ResNet50 feature 389 \nextractor was adapted into a Siamese network configuration, as conceptually illustrated in Figure 390 \n10. 391 \nThis architecture consists of two identical subnetworks, each acting as a weight-sharing \"finger 392 \nsequence network.\" Each subnetwork is modified to produce a 256-dimensional embedding 393 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n20 \n \nvector for its respective input image. The primary function of these embeddings is to serve as 394 \nlow-dimensional vector representations that encode the salient features of the image, thus 395 \nreducing data dimensionality while preserving discriminative information. Following the feature 396 \nextraction, the two embedding vectors are passed to the Euclidean Distance layer. This metric 397 \nserves as a distance measure, quantifying the relative dissimilarity between the two inputs by 398 \ncalculating the L2 norm between the two 256-dimensional vectors. This distance value is then 399 \nfed into a sigmoid function. The sigmoid output produces a probability of matching ranging 400 \nbetween 0 and 1, where a value near 0 indicates high similarity (the images belong to the same 401 \nclass) and a value near 1 indicates low similarity (the images belong to different classes or 402 \nclasses). 403 \n 404 \n 405 \n 406 \nFigure 8. An initial architecture 407 \n 408 \nThe initial network architecture relied on a 20-dimensional output vector where each element 409 \nwas mapped to a specific classification class. In contrast, the Siamese network adopts a metric 410 \nlearning paradigm and is trained using input pairs, as depicted in Figure 11. A representative 411 \nimage, referred to as the anchor, is fed into one side of the network. The network's training 412 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n21 \n \nobjective is to learn an embedding space where the genuine pair (an image of the same class as 413 \nthe anchor) yields a distance close to 0, and the impostor pair (an image that is not of the same 414 \nclass as the anchor) yields a distance close to 1. Specifically, the network is trained with positive 415 \npairs, where the second image belongs to the same class as the anchor, and negative pairs, where 416 \nthe second image belongs to a different class (Figure 12). This comparative training mechanism 417 \nforces the network to create highly discriminative embeddings that maximize the separation 418 \nbetween inter-class feature vectors while minimizing the distance between intra-class feature 419 \nvectors. 420 \n 421 \n 422 \n 423 \n 424 \n 425 \n 426 \n 427 \n 428 \n 429 \n 430 \n 431 \n 432 \n 433 \n 434 \n 435 \nFigure 9. Example of Positive Pair (Index and Index) with Negative Pair (Index and Thumb) 436 \n 437 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n22 \n \n 438 \nIn this study, a 3 × 4 full factorial experimental design was implemented to rigorously evaluate the 439 \nsensitivity of the Siamese network's performance to two key hyper-parameters: the embedding 440 \ndimension and the similarity threshold. This systematic approach allowed us to quantify the impact 441 \nof these components on the network's precision, conditional accuracy, unconditional accuracy, 442 \nconditional recall, unconditional recall, and the resulting uncertainty factor. Specifically, we tested 443 \nfour embedding dimensions (64, 128, 256, and 512) against three similarity thresholds (0.1, 0.2, and 444 \n0.5). The inclusion of both the conditional and unconditional variants of accuracy and recall was 445 \ncrucial to precisely determine the magnitude of the \"uncertain factor\", the percentage of images that 446 \nfall within the threshold boundary and require manual intervention. To ensure statistical rigor and 447 \nmitigate the effect of random initialization, each experimental condition was replicated 10 times. The 448 \nresulting mean values and standard deviations are comprehensively summarized in Figure 13 and 449 \nFigure 14. 450 \n 451 \n \n(a) \n \n(b) \n \n(c) \n  \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n23 \n \n 452 \nFigure 10: Impact of Embeddings on Different Metrics: (a) Conditional Accuracy, (b) 453 \nUnconditional Accuracy, (c) Precision, (d) Unconditional Recall, and (e) Conditional Recall. 454 \n 455 \n 456 \nFigure 11: Impact of Thresholds on Different Metrics: (a) Conditional Accuracy, (b) Precision, 457 \n(c) Unconditional Recall, (d) Unconditional Accuracy, (e) Conditional Recall. 458 \n 459 \n 460 \nThe analysis of the experimental results demonstrates the differential impact of the tested hyper-461 \nparameters. Varying the embedding dimensions (64, 128, 256, and 512) had no statistically 462 \nsignificant impact on classification accuracy or the other evaluated metrics, as evidenced across 463 \nFigure 2 (a-e). This suggests that a 64-dimensional embedding is sufficient for feature 464 \ndiscrimination within this specific metric space. In sharp contrast, modifying the similarity 465 \nthreshold (using subsets of 0.1, 0.2, and 0.5) yielded a significant and systematic change in 466 \nperformance. Narrowing the threshold (i.e., decreasing the value from 0.5 to 0.1) consistently 467 \nresulted in a substantial increase in conditional accuracy and recall (Figure 14 (a-e)). Crucially, 468 \n(d) (e) \n \n(a) \n \n(b) \n \n(c) \n \n(d) \n \n(e) \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n24 \n \nthis improvement simultaneously increased the uncertain factor from 0% (at a 0.5 threshold) to 469 \n24% (at a 0.1 threshold). The uncertain factor explicitly represents the number of image 470 \ncomparisons where the Euclidean distance falls within the range of [Threshold, 1 - Threshold]. 471 \nThese are predictions for which the network lacks sufficient confidence to assign a definitive 472 \npositive or negative classification, consequently requiring manual review. The impact of this 473 \nfactor is clearly visible when comparing the conditional versus unconditional metrics of accuracy 474 \nand recall (Figure 13 (a) and (e)), where the introduction of the uncertain images drastically 475 \ndepresses the unconditional values. This systematic dependency means that while smaller 476 \nthreshold values yield higher certainty in the classifications made, they dramatically increase the 477 \nvolume of images labeled as \"uncertain\" (Figure 15). This creates a direct operational trade-off: 478 \nachieving a high conditional classification accuracy (e.g., 98%) demands accepting an elevated 479 \nuncertain rate (e.g., 24%) (Table 2). Enhancement of fingerprint image using multiple filters has 480 \nalso been demonstrated [30]. For practical application in databases containing millions of 481 \nrecords, a 24% uncertainty level translates to an immense number of prints that must be 482 \nmanually inspected or passed through more sophisticated, yet time-consuming, secondary 483 \nclassification algorithms. Therefore, successful deployment of this proof-of-concept network 484 \nrequires careful optimization to balance the requirement for high classification rigor against the 485 \noperational feasibility of managing the resulting uncertainty volume. The combination of a high-486 \nquality benchmark database and an innovative assessment approach are considered to be a strong 487 \nfoundation for future research, contributing to the development of more reliable and secure 488 \nbiometric systems [31]. Fingerprint matching for noisy and distorted patterns using a Siamese 489 \nNetwork with ResNet50 and multihead attention has been reported [32]. FootprintNet, a Siamese 490 \nnetwork that utilizes pre-trained convolutional neural networks, specifically EfficientNet, 491 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n25 \n \nMobileNet, and ShuffleNet, to improve the robustness and accuracy of footprint recognition has 492 \nrecently been proposed [33] 493 \n 494 \n 495 \nFigure 12. Combined Histogram of Predicted Probabilities of Results. 496 \n 497 \n 498 \nTable 2. Summary of results of threshold, accuracy, uncertain and recall relationship 499 \nThreshold Accuracy Uncertain Recall/True \nPositive Rate \n0.5 92.5% 0% 92.9% \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n26 \n \n0.2 96.8% 14% 96.4% \n0.1 98.4% 24% 98.0% \n 500 \n 501 \nCONCLUSION 502 \nFingerprints are indispensable for identification in modern biometrics, yet the integrity of large-503 \nscale databases is continuously threatened by erroneous classification, a critical concern where a 504 \nsingle mislabeled image can compromise an entire archive. To address this, we successfully 505 \ndeveloped a Siamese network for the precise classification of fingerprints by finger type and 506 \ncollection methodology (flat versus rolled impressions). A central objective was to analyze the 507 \npractical utility of this tool by systematically testing the effect of various embedding dimensions 508 \nand similarity thresholds (specifically 0.5, 0.2, and 0.1) on performance metrics. Our primary 509 \nfinding reveals a crucial trade-off for operational deployment: using a lower similarity threshold 510 \nsignificantly boosts conditional classification accuracy (achieving up to 98%) and precision, but 511 \nthis comes at the expense of classifying a much larger percentage of images as \"uncertain\" (up to 512 \n24). This forced manual review of millions of uncertain records in large datasets highlights the 513 \nchallenge of balancing optimal statistical accuracy with practical operational efficiency and cost. 514 \nThis research delivers a robust proof-of-concept tool that moves beyond pure statistical accuracy 515 \nto provide a quantifiable, practical method for identifying and flagging data integrity errors. The 516 \nframework is modality-agnostic, making it suitable for deployment across face, iris, and other 517 \nbiometric data. Ultimately, this work contributes to establishing a standardized preprocessing 518 \nprocedure, a lightweight and efficient system that can be applied before any large biometric 519 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n27 \n \ndataset is finalized, ensuring the accuracy and legitimacy of identity records while concurrently 520 \nachieving substantial savings in time and operational costs. 521 \n 522 \n 523 \nAuthor Contributions 524 \nJMD and TIS conceived idea; TIS, design, conducted experiments, interpreted data and prepared 525 \nmanuscript; JMD and NMN, supervised, reviewed and improved the manuscript; All authors 526 \nagreed to submit the manuscript. 527 \n 528 \nAcknowledgements 529 \nTIS is thankful to the funding GRA provided by the JMD during this study. 530 \n 531 \nConflict of Interest Statement 532 \nThe authors declare no conflict on interests.  533 \nData Availability Statement 534 \nAll data are included in this manuscript. 535 \nORCID 536 \nTahsin Islam Sakif: 0000-0001-5378-4775 537 \nN. M. Nasrabadi: 0000-0001-8730-627X 538 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n28 \n \nJ M Dawson: 0000-0002-4539-7588 539 \n 540 \n 541 \n 542 \n 543 \nReferences 544 \n [1] J. Rohrlich, “Homeland Security will soon have biometric data on nearly 260 million 545 \npeople,” Quartz, Nov. 07, 2019. https://qz.com/1744400/dhs-expected-to-have-biometrics-on-546 \n260-million-people-by-2022 (accessed Mar. 21, 2023). 547 \n[2] U. Rao and V. Nair, “Aadhaar: Governing with Biometrics,” South Asia: Journal of South 548 \nAsian Studies , vol. 42, no. 3, pp. 469–481, May 2019, doi: 549 \nhttps://doi.org/10.1080/00856401.2019.1595343. 550 \n[3] J. Lynch, “HART: Homeland Security’s Massive New Database Will Include Face 551 \nRecognition, DNA, and Peoples’ ‘Non-Obvious Relationships,’” Electronic Frontier 552 \nFoundation, Jun. 07, 2018. https://www.eff.org/deeplinks/2018/06/hart-homeland-securitys-553 \nmassive-new-database-will-include-face-recognition-dna-and 554 \n [4] H. Rudolph, L. Moy, and A. Bedoya, “Airport Face Scans: An Investigation,” 555 \nwww.airportfacescans.com, Dec. 17, 2017. https://www.airportfacescans.com/ (accessed Mar. 556 \n21, 2023). 557 \n[5] P. Higgins, “Standards for the electronic submission of fingerprint cards to the FBI,” Journal 558 \nof Forenseic Identification, vol. 45, no. 4, pp. 409–418, 1995. 559 \n[6] K. E. Wertheim, “Human Factors in Large-Scale Biometric Systems: A Study of the Human 560 \nFactors Related to Errors in Semiautomatic Fingerprint Biometrics,” IEEE Systems Journal, vol. 561 \n4, no. 2, pp. 138–146, Jun. 2010, doi: https://doi.org/10.1109/jsyst.2010.2049878. 562 \n[7] “What Is Data Deduplication? | Benefits & Use Cases | NetApp,” www.netapp.com. 563 \nhttps://www.netapp.com/data-management/what-is-data-deduplication/ 564 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n29 \n \n[8] P. S. Sudhish, A. K. Jain, and K. Cao, “Adaptive fusion of biometric and biographic 565 \ninformation for identity de-duplication,” Pattern Recognition Letters, vol. 84, pp. 199–207, Dec. 566 \n2016, doi: https://doi.org/10.1016/j.patrec.2016.10.011. 567 \n[9] M. I. Jordan and T. M. Mitchell, “Machine learning: Trends, perspectives, and prospects,” 568 \nScience, vol. 349, no. 6245, pp. 255–260, Jul. 2015, doi: 569 \nhttps://doi.org/10.1126/science.aaa8415. 570 \n[10] H. Lin and A. Jain, “Classification of fingerprint images,” vol. 2, pp. 665–672, 1999, 571 \nAvailable: https://www.researchgate.net/profile/Lin-Hong-572 \n12/publication/2929466_Classification_of_Fingerprint_Images/links/54733a5a0cf216f8cfaeb8b7573 \n/Classification-of-Fingerprint-Images.pdf 574 \n[11] K. Fukushima, “A neural network for visual pattern recognition,” Computer, vol. 21, no. 3, 575 \npp. 65–75, Mar. 1988, doi: https://doi.org/10.1109/2.32. 576 \n[12] D. O. Hebb, The organization of behavior : a neuropsychological theory . New York: 577 \nRoutledge, Taylor & Francis Group, 2012. 578 \n[13] F. Rosenblatt, The Perceptron . Ithaca, New York: Cornell Aeronautical Laboratory, 1958, 579 \npp. 85–4601. 580 \n[14] Q. Zhang and H. Yan, “Fingerprint classification based on extraction and analysis of 581 \nsingularities and pseudo ridges,” Pattern Recognition , vol. 37, no. 11, pp. 2233–2243, Nov. 582 \n2004, doi: https://doi.org/10.1016/j.patcog.2003.12.020. 583 \n[15] D. Michelsanti, A.-D. Ene, Y. Guichi, R. Stef, K. Nasrollahi, and T. B. Moeslund, “Fast 584 \nFingerprint Classification with Deep Neural Networks,” Proceedings of the 12th International 585 \nJoint Conference on Computer Vision, Imaging and Computer Graphics Theory and 586 \nApplications, 2017, doi: https://doi.org/10.5220/0006116502020209. 587 \n [16] B. Pandya, G. Cosma, A. A. Alani, A. Taherkhani, V. Bharadi, and T. M. McGinnity, 588 \n“Fingerprint classification using a deep convolutional neural network,” 2018 4th International 589 \nConference on Information Management (ICIM) , pp. 86–91, May 2018, doi: 590 \nhttps://doi.org/10.1109/infoman.2018.8392815. 591 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n30 \n \n[17] R. A. Searston and J. M. Tangen, “The Emergence of Perceptual Expertise with Fingerprints 592 \nOver Time,” Journal of Applied Research in Memory and Cognition , vol. 6, no. 4, pp. 442–451, 593 \nDec. 2017, doi: https://doi.org/10.1016/j.jarmac.2017.08.006. 594 \n[18] N. K. Ratha, K. Karu, Shaoyun Chen and A. K. Jain, \"A real-time matching system for large 595 \nfingerprint databases,\" in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 596 \n18, no. 8, pp. 799-813, Aug. 1996, doi: 10.1109/34.531800. 597 \n[19] Y. LeCun, K. Kavukcuoglu, and C. Farabet, “Convolutional networks and applications in 598 \nvision,” Proceedings of 2010 IEEE International Symposium on Circuits and Systems , pp. 253–599 \n256, May 2010, doi: https://doi.org/10.1109/iscas.2010.5537907. 600 \n [20] K. Simonyan and A. Zisserman, “Very Deep Convolutional Networks for Large-Scale 601 \nImage Recognition,” arXiv.org, 2014. https://arxiv.org/abs/1409.1556 602 \n[21] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” 603 \n2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, Jun. 604 \n2016, doi: https://doi.org/10.1109/cvpr.2016.90. 605 \n [22] V. Meshram and K. Patil, “Border-Square net: a robust multi-grade fruit classification in 606 \nIoT smart agriculture using feature extraction based Deep Maxout network,” Multimedia Tools 607 \nand Applications, May 2022, doi: https://doi.org/10.1007/s11042-022-12855-7. 608 \n [23] Q. Zheng, M. Zhang, and J. Yang, “A Bilinear Multi-Scale Convolutional Neural Network 609 \nfor Fine-grained Object Classification,” IAENG International Journal of Computer Science, vol. 610 \n45, no. 2, 2018, Available: https://www.iaeng.org/IJCS/issues_v45/issue_2/IJCS_45_2_12.pdf 611 \n[24] K. Sai Prasad and D. Pasupathy, “DEEP CONVOLUTIONAL NEURAL NETWORK 612 \nIMPLEMENTATIONS FOR EFFICIENT BRAIN STROKE DETECTION USING MRI 613 \nSCANS,” Journal of Theoretical and Applied Sciences , vol. 100, no. 17, 2022, Accessed: Apr. 614 \n24, 2023. [Online]. Available: http://www.jatit.org/volumes/Vol100No17/14Vol100No17.pdf 615 \n[25] J. Bromley et al. , “SIGNATURE VERIFICATION USING A ‘SIAMESE’ TIME DELAY 616 \nNEURAL NETWORK,” International Journal of Pattern Recognition and Artificial 617 \nIntelligence, vol. 07, no. 04, pp. 669–688, Aug. 1993, doi: 618 \nhttps://doi.org/10.1142/s0218001493000339. 619 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint \n\n31 \n \n[ 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 \nManifolds,” openaccess.thecvf.com, 2019. 621 \nhttps://openaccess.thecvf.com/content_ICCV_2019/html/Roy_Siamese_Networks_The_Tale_of622 \n_Two_Manifolds_ICCV_2019_paper.html (accessed Apr. 24, 2023). 623 \n[27] Google, “Classification: Accuracy  |  Machine Learning Crash Course,” Google Developers, 624 \n2019. https://developers.google.com/machine-learning/crash-course/classification/accuracy 625 \n[28] Google, “Classification: Precision and Recall  |  Machine Learning Crash Course 626 \n       |  Google Developers,” Google Developers , Mar. 05, 2019. 627 \nhttps://developers.google.com/machine-learning/crash-course/classification/precision-and-recall 628 \n[29] J. Korstanje, “The F1 score,” Medium, Aug. 31, 2021. https://towardsdatascience.com/the-629 \nf1-score-bec2bbc38aa6 630 \n[30] Shams H, Jan T, Khalil AA, Ahmad N, Munir A, Khalil RA. 2023. Fingerprint image 631 \nenhancement using multiple filters. PeerJ Computer Science 9:e1183 632 \nhttps://doi.org/10.7717/peerj-cs.1183 633 \n[31] Bouhamed, S.A. Exploring possibilistic fingerprint image quality analysis: a soft computing 634 \napproach for biometric database management. Int. J. Mach. Learn. & Cyber.  (2025). 635 \nhttps://doi.org/10.1007/s13042-025-02587-z  636 \n[32] T. Sawhney et al., Fingerprint matching for noisy and distorted patterns using a Siamese 637 \nNetwork with ResNet50 and multihead attenti on. IEEE Access, vol. 13, pp. 88047-88064, 2025, 638 \ndoi: 10.1109/ACCESS.2025.3567328. 639 \n[33] İ brahimoğ lu, N., Osmani, A., Ghaffari, A. et al. FootprintNet: a Siamese network method for 640 \nbiometric identification using footprints. J Supercomput 81, 714 (2025). 641 \nhttps://doi.org/10.1007/s11227-025-07170-5 642 \n 643 \n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted December 16, 2025. ; https://doi.org/10.64898/2025.12.14.694185doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}