• • We propose an efficient framework that leverages Conditional DCGAN and Conditional WGAN-GP) to produce accurate and discriminative finger vein representations, thereby enhancing the stability of the training process.

• • We evaluate proposed framework on two publicly available datasets. The datasets utilized, the topologies of the suggested C-DCGAN and C-WGAN-GP models, and the evaluation criteria used to examine the accuracy and dependability of the synthesized finger vein representations.

• • We conduct a comparative study employing transfer learning with popular CNN architectures, including VGG19, MobileNetV2, InceptionV3, and EfficientNetV2-M which are properly tuned and trained using both conventional data augmentation and GAN-based augmentation (C-DCGAN and C-WGAN-GP), to evaluate the effectiveness of the enhanced datasets.

• • Traditional hand-made approaches include maximum curvature (MC),¹⁴ Local binary patterns (LBP),¹⁶ based on Gabor filtering,¹⁵ and their variations. For instance, Yang et al.¹⁷ presented ASAVE algorithm that minimizes the impact of finger displacement by extracting vein networks using orientation map–guided curvature. Furthermore, Yang et al.¹⁸ introduced a weighted vein-coding and index-integration framework aimed at improving both recognition accuracy and computational efficiency. Local Binary Pattern (LBP) and its variants characterize local image structure by capturing relationships among neighboring pixels, which makes them relatively robust to variations such as finger displacement. For example, the Weighted Local Descriptor (WLD)¹⁹ operator, an extension of LBP, and the SVM-based approach²⁰ estimate vein width and approximate finger cross-sections using elliptical models, thereby improving feature extraction and helping to compensate for pose-related distortions in low-quality images. Similar to this, Lv et al.²⁶ presented a method for extracting finger vein features that combines the pyramid of histograms of oriented gradients (PHOG) with the binary pattern of phase congruency (BPPC) to enable multi-scale feature representation while lessening the effects of local disturbances like changes in illumination. The merged feature set is subjected to the kernel discriminative common vector (KDCV) approach²¹ in order to obtain more discriminative, non-linear representations with a KNN classifier for personal identification.

  • • In standard representation learning, PCA, LDA,²² and related methods are commonly used. For example, Qiu et al.²³ designed a double sliding-window method to precisely identify the locations of the phalangeal joints on finger vein images., followed by feature extraction using two-dimensional PCA(2DPCA) to achieve reliable recognition performance. More recently, Yang et al.⁷ proposed a low-rank representation constrained by low reconstruction error (SRLRRR), which uses low-rank structures to suppress noise and extract discriminative information, improving recognition accuracy.

  • • Using multilayer architectures, deep neural networks can automatically learn abstract and high-level representations of finger vein pictures.²⁴ For instance, Das et al.²⁵ created a five-layer deep CNN and showed how well it performed on four publicly available datasets. Shaheed and associates. Moreover, improper selection of SVM parameters²⁶ often leads to poor classification accuracy and unstable performance Convolutional neural networks (CNN)⁴ dominated computer vision research and were the most widely adopted architectures for finger vein recognition prior to the introduction of Vision Transformers (ViTs).²⁷²⁸ suggested a CNN with residual connections that is deeply separable, achieving significant improvements in recognition accuracy. Although these approaches have achieved remarkable performance, their reliance on highly complex deep neural architectures introduces significant computational and memory overhead, reducing their suitability for real-time and embedded applications.²⁹

  • • Lightweight deep learning architectures have attracted increasing attention for finger vein recognition as a means of addressing the computational limitations associated with conventional deep networks.³⁰ Shen et al.³¹ presented an effective matching framework optimized utilizing a triplet loss function to accommodate real-time biometric applications, achieving competitive recognition performance with lower computational complexity. Chai et al.³² looked into the factors that affect lightweight network design and found that semi-pre-trained lightweight models consistently outperform networks trained from scratch, while rectangular convolutional kernels offer better feature representation than traditional square kernels. Huang et al.³³ developed a frequency–spatial coupling network that simultaneously utilizes spatial and frequency-domain information, improving authentication accuracy by building on the complementing qualities of several feature domains.

• • By combining the concepts of affinity maturation and clonal selection, the Binary Multi Objective Clonal Algorithm (BMCA) effectively achieves multi-objective optimization while simulating the adaptive behaviour of the biological immune system. It concurrently maximizes entropy and minimizes Pearson correlation while encoding band subsets as binary strings.³⁹ By defining hyperspectral band selection, QUBO problem and using quantum superposition and tunnelling to avoid local minima, Quantum Annealer (QA)-based metaheuristics offer a novel method.⁴⁰ In a quantum-enhanced framework, simultaneous band selection and classification are made possible by integrated quantum classifiers such as QBoost.

Two sophisticated conditional GAN frameworks, Conditional WGAN-GP and Conditional DCGAN, were used to provide high-quality synthetic training data in order to make up for the small amount of real finger vein images available for each class. The architectural configuration of each model is discussed in detail in the following subsections.

The objective function in C-DCGAN is well crafted the limited availability of actual finger vein images per class was addressed using two state-of-the-art GAN architectures to produce superior training samples. Among them, Conditional Deep Convolutional GAN (C-DCGAN)³⁵ extends the conventional DCGAN framework by conditioning both the generator and discriminator on class labels, thereby enabling the generation of class-aware finger vein images.

Generator architecture: A 100-dimensional latent noise vector concatenated with one-hot encoded class labels is sent into the generator. Its four transposed convolutional layers are followed by ReLU activation and batch normalization. It ends with a Tanh-activated output layer that produces 180 × 50-pixel grayscale images.

Discriminator architecture: Five convolutional layers with Leaky ReLU activation (α = 0.2) make up the discriminator. All layers—aside from the first and last layers—are subject to batch normalization. Real and synthetic images are distinguished using a sigmoid-activated output layer after class label embeddings are merged at the feature level.

In a conditional GAN architecture, the auxiliary information y, including multimodal attributes or class labels, conditions the generator G (z|y) and the discriminator D (x|y), which are integrated as an additional input to direct the data creation process. In the generator a joint hidden representation is created by combining the conditional variable y with the prior noise input p_z(z). This hidden representation may be built with a great deal of flexibility because of the adversarial training architecture.The objective function in C-DCGAN is well crafted to imitate the two-player mini-max game:

G D

The generator transmits the input noise vector z ∼ p_z(z) and conditional information y, while the discriminator evaluates the authenticity of the generated and real samples conditioned on y to synthetic samples. The Non-saturating Generator Loss and Discriminator Loss functions are given below.

By using Wasserstein loss in combination with regularization of the gradient penalty, C-WGAN-GP enhances training stability and sample variety.³⁶ The generator architecture follows that of the C-DCGAN, with the Wasserstein objective in place of the binary cross-entropy loss. The architecture of Conditional WGAN-GP is explained in Figure 2, which also demonstrates how Generator G and Critic D are used to train and validate both the produced and real samples. The discriminator, or critic, adds a gradient penalty term with a regularization coefficient of λ = 10 and does away with sigmoid activation. The generator loss is

The definition of a Wasserstein loss is:

where xˆ is evenly sampled between real and produced samples along straight lines.

Hyperparameter Selection: Optimal hyperparameters for both architectures were established through ablation studies. The latent space dimensionality was set to 100, providing a favorable balance between generation quality and training constancy. For C-DCGAN and Conditional WGAN-GP, respectively, learning rates of 0.0002 and 0.0001 were used to provide steady critic optimization and efficient generator learning. Adam optimizer was used for both models with β1 = 0.5 and β2 = 0.999.

Training Protocol: With a batch size of 64, each Conditional DCGAN and Conditional WGAN-GP were trained for 100 and 50 epochs, respectively. Observing (i) the convergence behavior of the discriminator and generator losses and (ii) a qualitative visual examination of the synthesized samples to verify.

anatomical plausibility was used to evaluate training stability. Table 1 lists the hyperparameter setup settings for the Conditional DCGAN and Conditional WGAN-GP implementation.

The VGG19 architecture was modified to process 180 × 50 × 1 grayscale pictures of finger veins. Five consecutive convolutional blocks with maximum-pooling layers 2 × 2 come after each of the 3 × 3 kernels with ReLU activation. The convolutional feature extractor is followed by Each of the three completely connected layers having 4096, 4096, and output-class neurons. To reduce overfitting, fully linked layers were subjected to a dropout rate of 0.5, and batch normalization after convolutional layers improved the stability and convergence of the training.

MobileNetV2 adopts a lightweight design based on Separable convolutions according to depth and inverted residual blocks with linear bottlenecks. The network was fine-tuned to accept 180×50×1 grayscale input by modifying the initial convolutional layer. Batch normalization and ReLU6 activation were utilized throughout the network to enhance numerical stability. Before the final classification layer, to enhance generalization and reduce parameter complexity, a global average pooling layer was employed.

InceptionV3 uses parallel convolutional filters with different receptive fields inside Inception modules to extract multi-scale features. By altering the input layer, the architecture was modified for grayscale photographs of the finger veins with a resolution of 180 × 50 × 1. In order to reduce computational cost and preserve effective gradient propagation during training, the design kept factorized convolutional layers. Before being transmitted to the last fully connected layer, whose output neurons match the predetermined target classes, the resultant feature representations were compressed using global average pooling.

In order to improve image quality and guarantee data consistency, the suggested system starts with a thorough preprocessing pipeline applied to the raw finger vein images. This pipeline includes region of interest (ROI) extraction, Contrast Limited Adaptive Histogram Equalization (CLAHE), and image normalization. A GAN-based augmentation module made up of Conditional Deep Convolutional GAN (C-DCGAN) and Conditional Wasserstein GAN with Gradient Penalty (C-WGAN-GP) receives the pre-processed pictures. Within this module, the generator synthesizes high-fidelity, identity-preserving finger vein images, while the discriminator learns to identify synthetic samples from authentic ones, therefore boosting the variety and representativeness of the training dataset. A convolutional neural network (CNN) based on transfer learning is then trained using the expanded dataset to carry out the final finger vein classification. Algorithm 1 provides a thorough explanation of the suggested methodology, and Figure 3 depicts the general structure.

The region of interest (ROI) photos in the THU-FVFDT2 database⁵ include information on dorsal finger texture and finger veins that were gathered from 610 different patients. Of these, the first 220 participants come from the previous THU-FVFDT1 dataset, while the remaining 390 subjects were recruited and scanned over the course of two capture sessions spaced roughly three to seven days apart. The dataset is ideal for assessing the robustness and generalization capacity of finger vein recognition algorithms since this multi-session acquisition procedure creates genuine temporal changes. One session is set aside for testing and the other for model training. Every ROI has a 200 × 100-pixel resolution adjusted. Most of the participants are volunteers from the Graduate School of Tsinghua University in Shenzhen, including students and staff.

The FV-USM ³⁷ database was created using photos obtained from 123 participants, comprising 40 female and 83 male University of Sains Malaysia employees and students. The participants were between the ages of 20 and 52. Every person’s left index, left middle, right index, and right middle fingers were recorded. The finger and vein patterns are two important biometric features in the obtained images. During a single acquisition session, each finger was scanned six times., which took place over two weeks apart for each participant. In the first session, 2,952 photos (123 × 4 × 6) were collected. Consequently, 5,904 pictures from both sessions—or 492 finger lessons—are included in the collection. A depth resolution of 256 gray levels and a spatial resolution of 640 × 480 pixels were used to capture finger images.

A personal computer with 16 GB of system memory and an Intel i5-8300H CPU operating at 2.30 GHz, and an NVIDIA GTX 1050 graphics processing unit with 8 GB of dedicated GPU RAM were used for all trials. The PyTorch framework (version 2.7), the torchvision library (version 0.22), and several auxiliary Python libraries for data preprocessing, training, and assessment were used to create deep learning models and related algorithms. GPU acceleration was used to boost processing efficiency during model training and inference.

The widely utilized approach of Contrast Limited Adaptive Histogram Equalization³⁸ was utilized to enhance local contrast. While preserving normalized general brightness to improve image quality. Figure 4. shows a sample of the original and enhanced images from the THU-FVFDT2 data set. Histogram equalization is applied inside each sector once the image is divided into tiny, non-overlapping tiles, the method improves local contrast while limiting contrast amplification using a preset clip limit. In this method, the grayscale photos are processed using CLAHE with an 8x8 tile size and a clip limit of 3.0. This preprocessing procedure increases the vein pattern’s visibility, reduces illumination fluctuations, and inhibits noise amplification, all of which increase the dependability of later Phases of feature extraction and classification.

For the THU-FVFDT2 and FV-USM datasets, at first, the training set was expanded with additional data. The augmentation process involved manually cropping the finger vein region, rotating the image within a ±10° range, converting it to grayscale, and resizing it to 50 × 180 pixels (width × height). In previous research, augmentation was performed using state-of-the-art techniques for image improvement and traditional geometric alterations.

In the suggested methodology, Figure 5 illustrates sample augmented images from the THU-FVFDT2 and FV-USM datasets generated using a Conditional WGAN-GP model. Furthermore, Figure 6 presents the generator and discriminator loss curves over 50 training epochs for the THU-FVFDT2 and FV-USM datasets, respectively, obtained using Conditional WGAN based augmentation methods.

This compares various deep learning architectures for finger vein detection using datasets enriched with the conditional DCGAN. The performance of the VGG19, MobileNetV2, InceptionV3, and EfficientNetV2-M models³⁹ is evaluated to determine how the extension of the C-DCGAN-based data affects the classification efficacy. The comparison findings demonstrate the resilience and adaptability of the suggested C-DCGAN-assisted framework to various topologies of convolutional neural networks.

Table 4 and 5 includes the performance of the various CNN models from FV-USM data set for the C-DCGAN model, and EfficientNetV2-M and MobileNetV2 perform better with dividing of 70–15-15 and 80–10-10,⁴⁰ respectively. Figures 7 and 8 show the graphs of Loss over epoch, Accuracy over epoch, and Macro Average Graph of the EfficientNetV2-M deep learning model.

Tables 6 and 7 include the performance of various CNN models from THU-FVFDT2 data set for the C-DCGAN model, and EfficientNetV2-M and MobileNetV2 perform better with the dividing of 70–15-15 and 80–10-10, respectively. The loss over epoch, accuracy over epoch, and macro average graphs of the MobileNetV2 deep learning model are displayed in Figures 9 and 10.

Figure 11 describes performance metrics using the conditional DCGAN augmentation technique with four transfer learning techniques with a data split of 70–15-15 and 80–10-10, respectively.

This paper uses datasets enhanced with a Conditional WGAN-GP to compare several CNN architectures. The impact of gradually enhanced Conditional WGAN-GP data on classification performance is investigated by analyzing the performance of VGG19, MobileNetV2, InceptionV3, and EfficientNetV2-M. The experimental findings show that the suggested Conditional WGAN-GP–based frame-work is robust and has better generalization capabilities across various convolutional neural network designs.

Tables 8 and 9 present the performance of various deep learning models that were trained using the C-WGAN-GP framework on the FV-USM data set. VGG19 and EfficientNetV2, two of the topologies assessed, perform better under the data partitioning techniques 70–15-15 and 80–10-10, respectively. The VGG19 model’s training dynamics are depicted in Figures 12 and 13, together with the macro-average performance graph, accuracy against the epoch curve, and loss vs epoch curve.

Tables 10 and 11 present the performance results of different deep learning models assessed using the THU- FVFDT2 dataset using the Conditional WGAN-GP frame-work. Among these, EfficientNetV2-M achieved superior performance with a data split of 70–15-15 and 80–10-10, respectively.

Figure 14 presents the comparative performance in terms of precision, precision, and recall using the C-WGAN-GP conditional enhancement framework. The evaluation is conducted with four transfer learning models—VGG19, MobileNetV2, InceptionV3, and EfficientNetV2-M—under data split configurations of 70–15–15 and 80–10–10, respectively.

The proposed approach uses Conditional DCGAN(C-DCGAN) and Conditional WGAN-GP(C-WGAN-GP) for both datasets, and C-DCGAN performs better than C-WGAN-GP. EfficientNetV2-M and MobileNetV2 were used as classifiers to evaluate the impact of augmented data. The results demonstrate consistently strong performance across both datasets, with maximum test accuracy of 0.9985 achieved for FV-USM database using EfficientNetV2-M with the 70–15–15% split. In general, these findings confirm that the integration of C-DCGAN-based augmentation with an appropriate data partitioning strategy effectively enhances finger vein recognition accuracy while maintaining robust generalization capability.

The variation in precision between training epochs for several C-DCGAN-based fin-ger vein identification models assessed on the THU-FVFDT2 and FV-USM datasets is shown in Figure 15. Effective feature learning made possible by GAN-driven data augmentation is demonstrated by all techniques’ quick accuracy gains in the early epochs, which are followed by stable convergence. In the FV-USM data set, EfficientNetV2-M performs best with a 70–15 to 15 splits, with an accuracy of 0.9985. Additionally, MobileNetV2 shows a competitive accuracy of more than 0.997 with reduced computational cost. The training stability and regularization advantages of the conditional DCGAN architecture are confirmed by smooth convergence trends.

A comparison of the suggested C-DCGAN and current finger vein detection techniques in several reference datasets is shown in Table 12. In both the THU-FVFDT2 and FV-USM datasets, the suggested method achieves the highest recognition accuracy, demonstrating its superior performance over previous techniques.

The datasets analysed in this study are not owned by the authors and therefore cannot be redistributed through this publication. The FV-USM and THU-FVFDT2 finger vein datasets are subject to the respective data providers' terms of use and distribution policies.

Researchers wishing to obtain these datasets should contact the original dataset providers and comply with their access requirements. The datasets are available only for research purposes and may require completion of a data access agreement or institutional approval before access is granted.

No additional ethical restrictions were imposed by the authors' institution because this study used previously collected and anonymized public research datasets. The Institutional Review Board (IRB) of the authors' institution did not require additional approval for data sharing, as no human participants were directly recruited by the authors and no identifiable personal information was collected.

The THU-FVFDT2 dataset is available at Graduate School at Shenzhen, Tsinghua University, upon reasonable request by submitting a signed license agreement to [email protected]. After the request is approved, the dataset can be downloaded, and the password required to extract the archived files will be provided by the dataset administrator.

The FV-USM dataset is available at University of Sains Malaysia upon request. Researchers may download the dataset from the official repository and obtain the password for accessing the archived files by contacting [email protected].

For accessing to data, the reviewers or readers may contact to corresponding author with the email address: [email protected]

The authors sincerely thank the Kalinga Institute of Industrial Technology, Bhubaneswar, for providing the necessary support and infrastructure that facilitated the publication of this article.

• 1.  Kumar A, Zhou Y: Human identification using finger images. IEEE Trans. Image Process. 2011; 21(4): 2228–2244. PubMed Abstract | Publisher Full Text
• 2.  Kim SG, Choi J, Hong JS, et al.: Spoof detection based on score fusion using ensemble networks robust against adversarial attacks of fake finger-vein images. Journal of King Saud University-Computer and Information Sciences. 2022; 34(10): 9343–9362. Publisher Full Text
• 3.  Qin H, Xi H, Li Y, et al.: Adversarial learning-based data augmentation for palm-vein identification. IEEE Trans. Circuits Syst. Video Technol. 2023; 34(6): 4325–4341. Publisher Full Text
• 4.  Zhang Z, Zhou Z, Yang X, et al.: Convolutional neural network based on multi-directional local coding for finger vein recognition. Inf. Sci. 2023; 623: 633–647. Publisher Full Text
• 5.  Yang W, Hui C, Chen Z, et al.: Fv-gan: Finger vein representation using generative adversarial networks. IEEE Trans. Inf. Forensics Secur. 2019; 14(9): 2512–2524. Publisher Full Text
• 6.  Choi J, Noh KJ, Cho SW, et al.: Modified conditional generative adversarial network-based optical blur restoration for finger-vein recognition. IEEE Access. 2020; 8: 16281–16301. Publisher Full Text
• 7.  Ou W-F, Po L-M, Zhou C, et al.:
• 8.  Fusion loss and inter-class data augmentation for deep finger vein feature learning. Expert Syst. Appl. 2021; 171: 114584. Publisher Full Text
• 9.  Tran N-T, Tran V-H, Nguyen N-B, et al.: On data augmentation for gan training. IEEE Trans. Image Process. 2021; 30: 1882–1897. Publisher Full Text
• 10.  Zhao P, Song Y, Wang S, et al.: Vpc-former: A transformer-based multi-view finger vein recognition model and a new benchmark. Pattern Recogn. 2023; 148: 110170. Publisher Full Text
• 11.  Hong HG, Lee MB, Park KR: Convolutional neural network-based finger-vein recognition using nir image sensors. Sensors. 2017; 17(6): 1297. Publisher Full Text
• 12.  Qin H, Qin L, Xue L, et al.: Finger-vein verification based on multi-features fusion. Sensors. 2013; 13(11): 15048–15067. PubMed Abstract | Publisher Full Text | Free Full Text
• 13.  Yang W, Qin C, Liao Q: A database with roi extraction for studying fusion of finger vein and finger dorsal texture. Chinese Conference on Biometric Recognition. Springer; 2014; pp. 266–270. Publisher Full Text
• 14.  Miura N, Nagasaka A, Miyatake T: Extraction of finger-vein patterns using maximum curvature points in image profiles. IEICE Trans. Inf. Syst. 2007; E90-D(8): 1185–1194. Publisher Full Text
• 15.  Peng J, Wang N, Abd El-Latif AA, et al.: Finger-vein verification using gabor filter and sift feature matching. 2012 Eighth International Conference on Intelligent Information Hiding and Multimedia Signal Processing. IEEE; 2012; pp. 45–48.
• 16.  Rosdi BA, Shing CW, Suandi SA: Finger vein recognition using local line binary pattern. Sensors. 2011; 11(12): 11357–11371. PubMed Abstract | Publisher Full Text | Free Full Text
• 17.  Yang L, Yang G, Yin Y, et al.: Finger vein recognition with anatomy structure analysis. IEEE Trans. Circuits Syst. Video Technol. 2017; 28(8): 1892–1905. Publisher Full Text
• 18.  Yang L, Yang G, Xi X, et al.: Finger vein code: From indexing to matching. IEEE Trans. Inf. Forensics Secur. 2018; 14(5): 1210–1223. Publisher Full Text
• 19.  Huang B, Dai Y, Li R, et al.: Finger-vein authentication based on wide line detector and pattern normalization. 2010 20th International Conference on Pattern Recognition. IEEE; 2010; pp. 1269–1272.
• 20.  Lee HC, Kang BJ, Lee EC, et al.: Finger vein recognition using weighted local binary pattern code based on a support vector machine. Journal of Zhejiang University SCIENCE C. 2010; 11(7): 514–524. Publisher Full Text
• 21.  Sarangi PP, Nayak DR, Panda M, et al.: A feature-level fusion based improved multimodal biometric recognition system using ear and profile face. J. Ambient. Intell. Humaniz. Comput. 2022; 13(4): 1867–1898. Publisher Full Text
• 22.  Zhao X, Guo J, Nie F, et al.: Joint principal component and discriminant analysis for dimensionality reduction. IEEE Trans. Neural Networks Learn. Syst. 2019; 31(2): 433–444. Publisher Full Text
• 23.  Qiu S, Liu Y, Zhou Y, et al.: Finger-vein recognition based on dual-sliding window localization and pseudo-elliptical transformer. Expert Syst. Appl. 2016; 64: 618–632. Publisher Full Text
• 24.  Wang G, Sun C, Sowmya A: Multi-weighted co-occurrence descriptor encoding for vein recognition. IEEE Trans. Inf. Forensics Secur. 2019; 15: 375–390. Publisher Full Text
• 25.  Das R, Piciucco E, Maiorana E, et al.: Convolutional neural network for finger-vein-based biometric identification. IEEE Trans. Inf. Forensics Secur. 2018; 14(2): 360–373. Publisher Full Text
• 26.  Kapoor K, Rani S, Kumar M, et al.: Hybrid local phase quantization and grey wolf optimization based svm for finger vein recognition. Multimed. Tools Appl. 2021; 80(10): 15233–15271. Publisher Full Text
• 27.  Li X, Zhang B-B: Fv-vit: Vision transformer for finger vein recognition. IEEE Access. 2023; 11: 75451–75461. Publisher Full Text
• 28.  Shaheed K, Mao A, Qureshi I, et al.: Ds-cnn: A pre-trained xception model based on depth-wise separable convolutional neural network for finger vein recognition. Expert Syst. Appl. 2022; 191: 116288. Publisher Full Text
• 29.  Xie C, Kumar A: Finger vein identification using convolutional neural network and supervised discrete hashing. Deep Learning for Biometrics. Springer; 2017; pp. 109–132.
• 30.  Zhao D, Ma H, Yang Z, et al.: Finger vein recognition based on lightweight cnn combining center loss and dynamic regularization. Infrared Phys. Technol. 2020; 105: 103221. Publisher Full Text
• 31.  Shen J, Liu N, Xu C, et al.: Finger vein recognition algorithm based on lightweight deep convolutional neural network. IEEE Trans. Instrum. Meas. 2021; 71: 1–13.
• 32.  Chai T, Li J, Prasad S, et al.: Shape-driven lightweight cnn for finger-vein biometrics. J. Inf. Secur. Appl. 2022; 67: 103211. Publisher Full Text
• 33.  Huang J, Zheng A, Shakeel MS, et al.: Fvfsnet: Frequency-spatial coupling network for finger vein authentication. IEEE Trans. Inf. Forensics Secur. 2023; 18: 1322–1334. Publisher Full Text
• 34.  Bowles C, Chen L, Guerrero R, et al.: Gan augmentation: Augmenting training data using generative adversarial networks. arXiv preprint arXiv:1810.10863. 2018.
• 35.  Mirza M, Osindero S: Conditional Generative Adversarial Nets.2014. Reference Source
• 36.  Jenkins J, Roy K: Exploring deep convolutional generative adversarial networks (dcgan) in biometric systems: a survey study. Discov. Artif. Intell. 2024; 4(1): 42. Publisher Full Text
• 37.  Asaari MSM, Suandi SA, Rosdi BA: Fusion of band limited phase only correlation and width centroid contour distance for finger-based biometrics. Expert Syst. Appl. 2014; 41(7): 3367–3382. Publisher Full Text
• 38.  Wanto A, Yuhandri Y, Okfalisa O: Optimization accuracy of cnn model by utilizing clahe parameters in image classification problems. 2023 International Conference on Networking, Electrical Engineering, Computer Science, and Technology (IConNECT). IEEE; 2023; pp. 253–258.
• 39.  Mathew A, Amudha P: Improved finger vein recognition using generative adversarial network and transfer learning. J. Nanoelectron. Optoelectron. 2024; 19(10): 1042–1052. Publisher Full Text
• 40.  Sumalatha U, Prakasha KK, Prabhu S, et al.: Enhancing finger vein recognition with image preprocessing techniques and deep learning models. IEEE Access. 2024; 12: 173418–173440. Publisher Full Text

Current Reviewer Status:

AWAITING PEER REVIEW

AWAITING PEER REVIEW

?