ISSN (Online) 2581-9429

IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 3, Issue 2, March 2023

Impact Factor: 7.301

Review Paper on an Authentication System using

Siamese Convolutional Neural Networks
Shrey Kekade1, Piyush Morey2, Mayur Rajput3, Sahil Karli4, Priyanka Bendale5
Students, Department of Computer Engineering1,2,3,4
Guide, Department of Computer Engineering5
Sinhgad College of Engineering, Pune, Maharashtra, India

Abstract: Due to its distinct advantages, finger vein verification has lately drawn more attention. Focusing
on the characteristics of finger vein verification, construct a Siamese structure combining with a modified
contrastive loss function for training the above CNN, which effectively improves the network's performance.
The experimental findings demonstrate that the lightweight CNN's size shrinks to 1/6th of the pretrained-
weights based CNN and that it achieves an equal error rate of 75% in the SDUMLA-HMT dataset, which
outperforms cutting-edge techniques and nearly maintains the same performance as CNN that is based on
pretrained weights.

Keywords: Finger Vein Verification, Convolutional Neural Network, Siamese CNN

I. INTRODUCTION
Image acquisition, image pre-processing, feature extraction, and matching are the four main components of a finger
vein verification system, with feature extraction serving as the most crucial stage. The system's total performance can be
markedly enhanced with discriminative image features. Three groups can be made based on the various feature
extraction techniques: finger vein pattern-based techniques, finger vein texture-based techniques, and minutiae-based
techniques.
In order to solve the image matching issue of signature verification, Siamese nets were first developed by Bromley and
LeCun in the early 1990s. (Bromley et al., 1993). Twin networks that take different inputs but are connected by an
energy function at the top make up a Siamese neural network. The highest-level feature representation on each side is
compared using some measure by this function. There is a connection between the twin networks' characteristics. The
same metric as if the opposite twins had been shown the same two pictures.
Weight tying ensures that because each network computes the same function, two extremely similar images cannot
possibly be mapped by their separate networks to extremely different locations in feature space. Additionally, because
of the network's symmetry, whenever we show two different images to the twin networks, the top conjoining layer
computes the same metric as if we had presented the same two images to the opposite twins.

II. LITERATURE SURVEY

Pattern-based methods: This kind of method mainly relies on the information provided by the global structure of the
blood vessels. Miura et al. [1] utilized repeated line tracking with randomly varied start points to extract the finger vein
patterns. Later, Liu et al. [2] modified the traditional line tracking algorithm to enhance its robustness and reduce the
computational cost and noise. In addition, Miura et al. [3] proposed a new method, which extracts the Centre-lines of
veins by finding local maximum curvatures in cross-sectional profiles of finger vein images with various widths and
brightness levels. Song et al. [4] also investigated the geometrical properties and extracted finger vein patterns using
their mean curvature. Moreover, Yang et al. [5] proposed a method to exploit vein information based on gabor filters.
Texture-based methods: Thetexture of a finger vein image is represented by a grey-scale distribution of pixels. Lee et al.
[6] used the local binary pattern (LBP) [7] and local derivative pattern (LDP) [8] to extract finger vein features. Later,
Rosdi et al. [9] proposed a new texture descriptor called the local line binary pattern, which shows better results than the
LBP and LDP. To further improve the performance,

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8874 501

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 3, Issue 2, March 2023

Impact Factor: 7.301

Yang et al. proposed the personalized best bit map [10], and personalized weight maps [11] which assign different
weight values for different bits according to their stability. Minutiae-based methods: The minutiae point in finger vein
verification refers to bifurcation points and endpoints. Yu et al. [12] extracted the minutiae features from vein patterns
and used a modified Hausdorff distance algorithm for matching. Liu et al. [13] proposed minutiae matching method
based on singular value decomposition. The scale- invariant feature transform (SIFT) features were exploited in [14, 15],
which show more tolerance to the rotation. In addition to the three main approaches described above, some other
methods have been investigated in recent years such as principal component analysis-based methods [16], manifold
learning [17] and superpixel-based methods [18]. In summary, the existing finger vein verification algorithms mainly
utilize the handcrafted features, which are sensitive to image quality and finger position. To overcome this issue, various
image preprocessing procedures and new features are also investigated, whose measures tend to be complex but the
improvements are limited.
Different from general images (such as that of a face), a finger vein image has its owncharacteristics. For example, they
are usually grey- scale images, and the information on texture features is relatively scarce and changes slowly and a
transition from texture to the background is not sufficiently obvious. Extracting more effective and discriminative
features from this type of image is especially vital to make the performances of finger vein verification systems more
accurate and robust.In view of the shortcomings within existing traditional finger vein verification algorithms based
on handcrafted features and the CNN-based finger vein verification methods, we propose a novel method based on the
Siamese CNN to better meet the requirements of finger vein verification. Its contributions are summarized as follows:
(i) through the use of image augmentation along with the method of building a CNN based on pretrained weights, we
alleviated the issues of lacking trainingsamples for deep learning. (ii) On the basis of the pretrained-weights based CNN,
the Siamese structure was constructed for metric learning, and we proposed a new modified contrastive loss (MC loss)
function for training, which further improved its discriminative power for features. (iii) Considering a deployment in
embedded devices with limited hardware resources, we first developed and trained a pretrained weights-based CNN,
whose knowledge was then transferred to a newly built lightweight CNN by a knowledge distillation method, which
made the final finger vein verification CNN model small but effective. (iv) We performed a sufficient number of open-
set experiments on three public datasets and one self- built dataset, which all verified that our method achieved a state-
of-the-art performance.

III. PROBLEM STATEMENT OF SYSTEM

Traditional authentication methods like passwords and magnetic cards are far from meeting the requirements of society
given the growing demand for accurate and reliable identity authentication in modern society. This is because they can
be easily stolen and falsified. Meanwhile, biometric identifiers (such as fingerprints, faces and iris patterns) are
characterized by their uniqueness and stability, making biometric recognition technology more important in the identity
authentication field. The finger veins are located beneath the skin, which makes them more difficult to steal or wear
than the majority of biometric identifiers that are on the exterior of the body. Most importantly, finger vein verification
has the special ability to identify live bodies, which guarantees the benefits of finger vein technology and draws more
interest to this field.

IV. IMPLEMENTATION DETAILS OF MODULE

The proposed system contains following:

4.1 Pre-processing
Data will be loaded into the system, checked for accuracy, and then trimmed and cleaned for analysis. Make sure to
thoroughly document the cleaning decisions and provide justification. The data that was gathered might have missing
numbers, which could cause it to be inconsistent. To gain better results data need to be pre-processed so as to improve the
efficiency of the algorithm. Variable conversion must be done after the anomalies have been removed.

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8874 502

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 3, Issue 2, March 2023

Impact Factor: 7.301

4.2 Building the Siamese CNN Model

The Siamese network with a contrastive loss function, which is fed sample pairs and extracts their features to learn the
similarity measure of the sample pairs, can meet the requirements for featuresin finger vein verification. Furthermore, the
Siamese network can use the learned similarity measure to compare and match samples of new unknown categories,
thus it is truly suitable for applying it to a task with a large number of categories, where the number of samples for each
category is very small after grouping the samples into pairs.

4.3 Working
Using Siamese neural networks, we learn image representations using a supervised metric-based approach, then reuse
their features for one-shot learning without retraining.
In spite of the fact that the basic approach can be applied to almost any modality, we focus on character recognition. To
tackle this problem, we use large Siamese convolutional neural networks, which are capable of:
1. Learn generic image features that can be used to predict unknown class distributions even when few examples
are available;
2. Could be easily trained on pairs of data samples using standard optimization techniques; and
3. Provide a competitive approach that does not rely on domain-specific knowledge by instead utilizing the
strengths of large Siamese convolutional neural networks.

V. CONCLUSION
One advantage of the suggested training structure is that lightweight CNN performs almost as well as the pretrained-
weights based CNN while using considerably fewer resources. The size of the dataset required for authentication
process will significantly reduce. In order to achieve an even better identification rate, we intend to enhance the imaging
gadget and algorithm.
Additionally, we can attempt to create brand-new deep learning modules targeted at particular finger vascular
deformations

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8874 503

www.ijarsct.co.in
 ISSN (Online) 2581-9429
IJARSCT
International Journal of Advanced Research in Science, Communication and Technology (IJARSCT)

Volume 3, Issue 2, March 2023

Impact Factor: 7.301

REFERENCES
[1]. Shruti Jadon, IEEE Member, [email protected], An Overview of Deep Learning Architectures in Few-
Shot Learning Domain, arXiv:2008.06365v3 [cs.CV] 19 Aug 2020.
[2]. Su Tang1 , Shan Zhou1 , Wenxiong Kang1 , Qiuxia Wu2 , Feiqi Deng1, Finger vein verification using a
Siamese CNN, 23rd May 2019
[3]. Miura, N., Nagasaka, A., Miyatake, T.: ‘Extraction of finger-vein patterns using maximum curvature points
in image profiles’, IEICE Trans. Inf. Syst., 2007, 90, (8), pp. 1185–1194
[4]. Song, W., Kim, T., Kim, H.C., et al.: ‘A finger- vein verification system using mean curvature’, Pattern
Recognit. Lett., 2011, 32, (11), pp. 1541–1547
[5]. Yang, J., Yang, J., Shi, Y.: ‘Finger-vein segmentation based on multi-channel even-symmetric gabor filters’.
IEEE Int. Conf. Intelligent Computing & Intelligent Systems, Shanghai, China, 2009, vol. 4, pp. 500–503
[6]. Lee, E.C., Jung, H., Kim, D.: ‘New finger biometric method using near infrared imaging’,
[7]. Sensors, 2011, 11, (3), pp. 2319–2333
[8]. Ojala, T., Pietikainen, M., Maenpaa, T.: ‘Multiresolution gray-scale and rotation invariant texture
classification with local binary patterns’, IEEE Trans. Pattern Anal. Mach. Intell., 2002, 24, (7), pp. 971–987
[9]. Baochang, Z., Yongsheng, G., Sanqiang, Z., et al.: ‘Local derivative pattern versus local binary pattern: face
recognition with high-order local pattern descriptor’, IEEE Trans. Image Process., 2010, 19, (2), pp. 533–544
[10]. Rosdi, B.A., Shing, C.W., Suandi, S.A.: ‘Finger vein recognition using local line binary pattern’, Sensors,
2011, 11, (12), pp. 11357–11371
[11]. Yang, G., Xi, X., Yin, Y.: ‘Finger vein recognition based on a personalized best bit map’, Sensors, 2012, 12,
(2), pp. 1738–1757
[12]. Yang, G., Xiao, R., Yin, Y., et al.: ‘Finger vein recognition based on personalized weight maps’, Sensors,
2013, 13, (9), pp. 12093–12112
[13]. Yu, C.-B., Qin, H.-F., Cui, Y.-Z., et al.: ‘Finger-vein image recognition combining modified Hausdorff
distance with minutiae feature matching’, Interdiscip. Sci. Comput. Life Sci., 2009, 1, (4), pp. 280–289
[14]. Liu, F., Yang, G., Yin, Y., et al.: ‘Singular value decomposition based minutiae matching method for finger
vein recognition’, Neurocomputing, 2014, 145, pp. 75–89
[15]. Pang, S., Yin, Y., Yang, G., et al.: ‘Rotation invariant finger vein recognition’ (Springer Berlin Heidelberg,
Berlin, Heidelberg, 2012), pp. 151–156
[16]. Peng, J., Wang, N., El-Latif, A.A.A., et al.: ‘Finger-vein verification using gabor filter and SIFT feature
matching’. 2012 Eighth Int. Conf. Intelligent Information Hiding and Multimedia Signal Processing, Athens,
Greece, 18– 20 July 2012, pp. 45–48
[17]. Wu, J.-D., Liu, C.-T.: ‘Finger-vein pattern identification using principal component analysis and the neural
network technique’, Expert Syst. Appl., 2011, 38, (5), pp. 5423–5427
[18]. Liu, Z., Yin, Y., Wang, H., et al.: ‘Finger vein recognition with manifold learning’, J. Netw. Comput. Appl.,
2010, 33, (3), pp. 275–282
[19]. Liu, F., Yin, Y., Yang, G., et al.: ‘Finger vein recognition with superpixelbased features’. IEEE Int. Joint
Conf. Biometrics, Florida, USA, 29 September 2014–2 October 2014, pp. 1

Copyright to IJARSCT DOI: 10.48175/IJARSCT-8874 504

www.ijarsct.co.in