Format Template


 

 Vol. 6, No. 1 | January – June 2022 
 

 

 

SJCMS | P-ISSN: 2520-0755| E-ISSN: 2522-3003 | Vol. 6 No. 1 January – June 2022 

16 

Finger-Vein Image Dual Contrast Adjustment and 

Recognition Using 2D-CNN

Noroz Khan Baloch1, Saleem Ahmed2, Ramesh Kumar2, DM Saqib Bhatti3, Yawar 

Rehaman4 

Abstract: 

The Recommended steps improve the low contrast of the finger-vein image by employing 

dual contrast adaptive histogram equalization (DCLAHE) for visual attributes. The dual CLAHE 

improves image histogram and it is split up throughout the image pixels. The dataset containing 

finger-vein images is from the SDUMLA-HMT finger-vein database. Following the deployment 

of DCLAHE. The new version of the dataset is used to detect objects by employing an improved 

2D-CNN method. The learning of features in the proposed model is based on optimizing the 

values of a preprocessed dataset. The proposed model produces 95.35% accuracy. 

Keywords: adaptive histogram equalization; convolutional neural network. 

1 Introduction  

At present, the recognition systems are 
based on biometric technologies which can be 
used for various security purposes like 
computer access, banks, phones, ATMs. Such 
recognition approaches are known as finger-
vein, iris, fingerprint, iris, [1]-[3]. In recent 
years, the recognition system which got much 
attention is known as the finger-vein 
recognition system due to vein patterns under 
the skin are distinct makes it a suitable method 
to be used for human classification based on 
the biometrics method [3]. The previous 
approaches for finger-vein recognition are 
based on highlighting the lines from the image 
or it improves the quality of the image and then 
performing feature extraction [4]. The vein of 
the fingers is usually extracted by using near-

 
1Dept. of Electronics Engg. Dawood University of Engineering  and technology Karachi, 

Pakistan. 
2Dept. of Computer System Engg. Dawood University of Engineering and technology Karachi, 

Pakistan.. 
3Dept. of Telecommunication Engg. Dawood University of Engineering and Technology 

Karachi, Pakistan. 
4Dept. of Electronic Engg., NED UET Karachi, Pakistan. 

 

infrared (NIR) light and a charged coupled 
device (CCD) camera beneath the finger to 
capture veins is used for image acquisition [5]. 
The first step is to normalize data as a 
preprocessing requirement. Therefore, each 
image of 240×320×3 is transformed into single 
dimension grayscale image of 128×160×1 
before feeding into preprocessing unit and then 
forwarded to 2D CNN model [5]. Hence 
finger-vein area becomes darker on the palmer 
side due to CCD camera on the palmar side of 
the finger and also due to the penetration of 
NIR on the dorsal side [6]. 

There are several datasets to be used for the 
finger-vein detection method. We employed 
one of the most common databases known as 
SDUMLA-HMT [7] to analyze the feasibility 
of the proposed approach for finger vein 
recognition. Few of the known datasets are 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

17 

MMCBNU_6000, HKPU-FV, THU-FV 
UTFV, THU-FV [8]-[11]. There are several 
approaches used to extract the finger vein part 
from image background like Gabor filters 
(GF), local binary pattern (LBP), and local 
derivative pattern (LDP) [12]-[14]. However, 
these methods' performance is lower due to 
limitations in familiar scenarios such as 
rotation during the acquisition time of the 
finger vein. Therefore, for such reason the 
classification of finger-vein images is 
performed by employing Artificial Intelligence 
techniques known as convolutional neural 
network (CNN). CNN/ConvNet is deep 
learning (DL) method that has the advantage of 
more accurate and rapid results compared to 
traditional algorithms [1]. 

In our proposed method we employ Dual 
CLAHE [7] based preprocessed dataset and 
then fed them to CNN model which employs 
2D-CNN model for classification. Prior to 
classification, our proposed model uses Sobel 
edge detector and polygonal region of interest 
(ROI) methods to extract the background of 
the finger region. Therefore, we employ 
histogram equalization to enhance the 
performance of the proposed method. The 
histogram illustrates the grayscale distribution 
of pixels in an image and distributes grey 
values in a whole image [16]. In our proposed 
method we employ dual contrast limited 
adaptive histogram equalization (CLAHE) to 
increase grey values in an image. The output 
images of dual CLAHE are prearranged as 
classes concerning their folder names. Then 
these finger vein images are fed into the 
proposed 2D-CNN model. As compared to 
traditional approaches the 2D-CNN method 
extracts the finger-vein from image datasets by 
convolving different filters on a large scale of 
the dataset by learning features and 
recognizing them without extracting features 
from the background. It is observed that 2D 
CNN can produce results of finger vein 
detection rate reaching near 95%.  

The paper is organized as follows. Section 
2 contains the work done in the field of finger-
vein detection. Then the proposed model and 
relevant approaches such as background 
extraction, Dual CLAHE, and 2D CNN 

architecture are provided in section 3. In 
section 4, the simulation results and model and 
its parameters are provided and a conclusion is 
given in section 5. 

2 Related work on Finger-vein 
Detection 

Previous research on finger vein detection 
is based on feature extraction and 
preprocessing [1]-[4].  The Gabor filters are 
used in order to extract different shapes and 
directions to detect the finger-vein pattern [17]. 
In [12], the author used a local derivative 
pattern (LDP) [12] and a local binary pattern 
(LBP) [18] to extract finger–vein patterns. By 
using the Gabor filter, the SIFT features of 
finger vein patterns are matched in the finger-
vein classification method proposed in [19]-
[20]. The distinguishable features are 
generated by applying the principal component 
analysis (PCA) technique for feature 
extraction [21]. Other few notable techniques 
for finger-vein feature extraction are SVM, 
HOB, and sparse representation techniques 
[23]. 

Instead of above mentioned traditional 
schemes, researchers are putting efforts to 
utilizes deep learning for finger-vein detection 
for better performance in the finger-vein 
detection [1]-[2]. In [24], CNN based approach 
with four-layer employing fused convolutional 
sampling is used for finger vein detection 
where they propose to reduce the complexity 
by applying mentioned approach. In [25], a 
seven-layer CNN model with five convolution 
layers and two fully connected layers is 
proposed. The CNN with multimodal finger 
fusion method is proposed in [26]. They also 
worked on the uneven size for feature 
extraction by applying PCA based 
dimensionality reduction method. However, 
conventional methods suffer from a 
complexity-performance tradeoff. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

18 

3 The Proposed Finger Vein Detection 
Method 

3.1 Overview 

Figure 1 shows the proposed finger vein 
detection method contains six steps. First, we 
read the images from an available database 
used in our study SDUMLA-HMT [7], which 
is an open source dataset that can be used for 
study based on finger-vein images. Each image 
has a size of 240×320×3 is normalized in the 
second step and it became 128×160×1 size 
image as shown in Fig. 2(a & b), respectively. 
Fig. 2(c) contains the background separation of 
an image which is the third step of our 
proposed method. The polygonal ROI and 
Sobel edge detector are used for finger region 
extraction. Then the dual CLAHE is used for 
yielding dark vein ridges. In the last step, the 
images are passed through 2D-CNN model. 
Table 1 shows the 2D CNN architecture 
containing 24 layers. 

 

 

3.2  Dual CLAHE 

Due to the low contrast of the finger vein 
images, it becomes difficult to separate the 
finger vein from the background or other 
regions. To resolve this issue, we employed the 
dual CLAHE technique on the finger-vein 
image. In the dual CLAHE method, first, the 
finger-vein image is cropped from the dark 
region. After cropping the image, we employ 
dual CLAHE to enlighten the vein region twice 
as shown in Fig. 3. CLAHE is derived from 
Adaptive Histogram Equalization (AHE) [28]. 
AHE is employed on small areas of the 
contextual region in the image known as tiles 
2×2 on both CLAHE. With clipping limits of 
0.03 and 0.04 on the first and second, 

respectively having the CLAHE with the same 
exponential distribution [7] [27]. 

𝑁𝑎𝑣𝑔 =
𝑁𝑟𝑋×𝑁𝑟𝑌

𝑁𝑔𝑟𝑎𝑦
 () 

Where N_avg is the average number of 
pixels, and a number of pixels in the X & Y 
directions are shown as N_r X & N_r Y, 
respectively, and N_gray is the number of 
gray-scale pixels. 

The CL (clipped limit) is used to limit the 
values at a certain specified level in a grayscale 
having values darker which are recognized as 
finger-vein as given below: 

𝑁𝐶𝐿 = 𝑁𝑐𝑙𝑖𝑝 × 𝑁𝑎𝑣𝑔   () 

𝑁avggray =
𝑁Σclip

𝑁𝑔𝑟𝑎𝑦 
   () 

Where the real CL is N_CL, the actual CL is 
N_clip in the [0-1] range. If pixel number 
N_CL is lower than CL, then it is clipped. The 
NΣclip is the number of clipped pixels. Then 
the remaining average grey-level pixels are 
distributed as shown in (3) [7] [27]. 

3.3  2D-CNN Architecture 

The 2D-CNN architecture has twenty-four 
layers. Initially, the image is input in the first 
layer having 50×50×1 size. Then there is a 
total of five 2D convolutional layers which 
apply different filters on images. The proposed 
model also contains five ReLU layers and five 
batch normalization layers. The proposed 
architecture contains a total of 3 max-pooling 
layers. There are 2 layers of fully connected 
layers followed by 1 softmax layer and finally, 
the output layer is known as the classification 
layer. The proposed model achieves an 
accuracy of 95.35 %. 

      The finger vein dataset is convolved with 
filters when they are passed through 
convolution layers. Where the features are 
learned for the given images and a feature 
matrix is generated which is given in Fig. 4 

Fig. 1. Steps involved in the proposed 

scheme. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

19 

which shows the output of 1st convolutional 
layer. The five filters of size  4×4   are 
employed in the first convolutional layer with 
padding of [2 2 2 2] zeros from each direction. 
The other convolutional layers we employed 
10, 20, 30, 40 filter size on each second, third, 
fourth, and fifth convolutional layer, 
respectively, as shown in table (1).  

 

 

  
𝑦 = ∑ 𝓌c𝑁𝑐=1 ∗ 𝓍𝑐 + 𝑏 () 

 

 

In (4) the 𝜔𝑐  , 𝑥𝑐   and b are the weight in 
the 𝑐𝑡ℎ input channel, 2D input of the 𝑐𝑡ℎ 
channel filter, and filter bias, respectively. The 
bias factor can neglected by apply batch 
normalization having 𝛽– shift factor as 
depicted in Fig. 5. The output of the 
convolutional layer, the bias factor can change 
the sizes of channels. Therefore, batch 
normalization is employed for channels 
normalization for ReLU activation [28]. 

The output of batch normalization is fed to 
the activation function in the ReLU layer that 
introduces nonlinearity as depicted in Fig. 6 
[5]. In our proposed system, we employ the 
maximum value of pixels. Pixels are 
considered as zero if pixels have a value less 
than zero and if the value is greater than zero 
they are considered y. The nonlinear functions 
employed during the ReLU layer are the tanh 
and sigmoid activation functions. The ReLU is 
well known established approach and it is as 
given in (5) [1] [5]:  

 𝑓(𝑦) = max (0, 𝑦) () 

 
(a)  

 
(b)  

 
(c)  

 

 

 

 

 

Fig. 3. Dual CLAHE implementation. 

Fig. 5. Output from Batch Normalization 

Layer. 

 

Fig. 4.Output of the convolutional layer. 

Fig. 2. (a) Original image of size 

240×320×3, as of dataset SDUMLA-HMT 

(b) dimension reduced Image 128×160×1, 

(c) Region extracted image by applying 

polygonal ROI and Sobel edge detector. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

20 

 

 
The max-pooling layer will take the maximum 
values with a filter matrix of 2×2 which will 
only take the maximum number of values of 
channels. The stride of the max-pooling layer 
is [1 1] with padding 0. The max-pooling 
output is shown below in Fig. 7. 

      he fully connected layer we use two times 
t=f_c (x,w,b) this function of the fully 
connected layer, where x represents the input 
of this layer, where w represents the weight 
matrix, b is the bias, and t is the output [29] 
which is as mentioned in (6): 

 𝑡𝑖′ = ∑ 𝑤𝑖𝑖′ 𝑥𝑖𝑖 + 𝑏𝑖  () 

The second last layer of 2D CNN produces 
a probability distribution for every image 
location used to classify pixels into one of the 
initial labelled where the related object classes 

belong to. The classification output layer uses 
the cross-entropy loss of reciprocally absolute 
classes for multi-class classification problems 
[30]. 

 

 

4 Result 

In the proposed method we have divided 
the dataset as 80% for training and 20% for 
validation and testing of the model. The Adam 
optimizer-based train network is fed with the 
partitioned dataset in order to optimize the 
learned values and extract features in fed 
images from the proposed 2D-CNN model. In 
the proposed method, four images were 
selected for preprocessing from left-hand and 
right-hand folders and tagging them with the 
designations of the index, middle, and ring 
fingers sub-folders. During the training, the 
process epoch was set to 80, where an epoch is 
a full cycle with 12 iterations. In Fig. 10, the 
training process illustrates validation accuracy 
of 95.35% after completing the 960 iterations. 
The 2D-CNN model is trained and the update 
of weights is done by using the Adam 
optimizer. Table (II) shows the various 
parameter like iterations, epoch size, and 
iterations per epoch. It also contains the 
validation accuracy achieved during these 
parameters used in the model.  Fig. 11 shows 
the loss of model in which weight can be 
renewed and to decrease the loss following 
evaluation

TABLE I.  THE 2D CNN ARCHITECTURE FOR THE PROPOSED METHOD. 

Fig. 6. ReLU layer output. 

Fig. 7. ReLU Layer Graph. 

Fig. 8. Output Channels of Max-pooling 

layer. 

Fig. 9. Basic architecture of 2D CNN. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

21 

# Layers Name Input Size Kernel size Stride Padding Filters 

1 Image Input layer 50 × 50 × 1     
2 Convolution2dlayer 1  4 × 4 [1 1] [2 2 2 2] 5 

3 
Batch Normalization 

layer 1 

 

 
   

4 ReLulayer 1      
5 Max Pooling 1  2 × 2 [1 1] [0 0 0 0]  

6 Convolution2dlayer 2  5 × 5 [1 1] [2 2 2 2] 10 

7 
Batch Normalization 

layer 2 

 

 
   

8 ReLulayer 2      

9 Max Pooling 2  2 × 2 [1 1] [0 0 0 0]  
10 Convolution2dlayer 3  6 × 6 [1 1] [2 2 2 2] 20 

11 
Batch Normalization 

layer 3 

 

 
   

12 ReLulayer 3      
13 Max Pooling 3  2 × 2 [1 1] [0 0 0 0]  
14 Convolution2dlayer 4  7 × 7 [1 1] [2 2 2 2] 30 

15 
Batch Normalization 

layer 4 

 

 
   

16 ReLulayer 4      
17 maxPooling2dLayer  2 × 2 [1 1] [0 0 0 0]  

18 Convolution2dlayer 5  8 × 8 [1 1] [2 2 2 2] 40 

19 
Batch Normalization 

layer 5 

 

 
   

20 ReLulayer 5      

21 
Fully Connected 

Layer 1 

 

 
   

22 
Fully Connected 

Layer 2 

 

 
   

23 Softmax      

24 Classification Output      

TABLE II.  2D CNN TRAINING PARAMETERS 

# Parameter Values 

1 Validation accuracy 95.35% 

2 Epoch 80 

3 Iteration 960 

4 Iteration per epoch 12 

5 Validation Frequency  30 iteration 

 

The proposed model [31] results are 
compared with a twenty-layer 2D-CNN in 
which the features of finger-vein images are 
learned with a validation accuracy of 94.87%, 
as illustrated in Fig. 12–13 below;  

 

 

 

 

This space is intentionally left blank 
to adjust the images 

 

 

 

 

 

 

 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

22 

 

 

  

 

 

 

 

 

 

This space is intentionally left blank to adjust the images 

 

Fig. 10. Accuracy graph of 2D CNN using 24 layers. 

Fig. 11. Loss graph of 2D CNN using 24 layers. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

23 

 

 

 

 

Fig. 10 and 12 show that the validation 
accuracy of the proposed model using 24 
layered 2D CNN yields much better results 
with a validation accuracy of 95.35%, while 
the previous results have a validation accuracy 
of 94.87%. In this proposed model, we use four 
max-pooling layers with a kernel size of 2 × 2 
with a stride of 1, while the previous model 
used three max-pooling layers with a filter size 

of 3 × 3 and stride of 2. Fig. 11 and 13 
compare the losses between the proposed 
model and the previous model of 2D-CNN in 
which validation losses are minimized as 
compared to the previous 2D-CNN model in 
[31]. 

 

Fig. 12. 2D-CNN accuracy graph using 20 layers. 

Fig. 13. The loss graph 2D-CNN using 20 layers [31]. 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

24 

5 Conclusion 

1. We deployed a 2D CNN for finger-vein 
recognition in this study, which uses 
preprocessed dual CLAHE augmented image 
features and takes 2544 finger-vein images. To 
distinguish the background from the finger 
region, the pre-enhanced finger-vein images 
are run via the Sobel edge detector and 
polygonal ROI. The dataset is then run through 
2D CNN to recognize the features of finger-
vein images with distinct classes of folders like 
the middle, index, and ring folders. The 2D-
CNN uses 4 images from each. As a result, our 
model was able to learn the characteristics of 
each image as well as recognize each image 
within folders. The accuracy of this proposed 
model is also near 96%. 

REFERENCES 

[1] Liu, Wenjie, Weijun Li, Linjun Sun, Liping 
Zhang, and Peng Chen, "Finger vein 
recognition based on deep learning," In 12th 
IEEE Conference on Industrial Electronics 
and Applications (ICIEA), p. 205-210. IEEE, 
2017 

[2] Pham, Tuyen Danh, Young Ho Park, Dat Tien 
Nguyen, Seung Yong Kwon, and Kang 
Ryoung Park, "Nonintrusive finger-vein 
recognition system using NIR image sensor 
and accuracy analyses according to various 
factors," Sensors 15, no. 7 (2015): 16866-
16894. 

[3] Mulyono, David, and Horng Shi Jinn, "A 
study of finger vein biometric for personal 
identification," In 2008 International 
Symposium on Biometrics and Security 
Technologies, pp. 1-8. IEEE, 2008. 

[4] Hong, Hyung Gil, Min Beom Lee, and Kang 
Ryoung Park, "Convolutional neural network-
based finger-vein recognition using NIR 
image sensors," Sensors 17, no. 6 (2017): 
1297. 

[5] Das, Rig, Emanuela Piciucco, Emanuele 
Maiorana, and Patrizio Campisi, 
"Convolutional neural network for finger-
vein-based biometric identification," IEEE 
Transactions on Information Forensics and 
Security 14, no. 2 (2018): 360-373. 

[6] Shin, Kwang Yong, and Kang Ryoung Park, 
"Finger-vein image enhancement using a 
fuzzy-based fusion method with gabor and 
retinex filtering," Sensors 14, no. 2 (2014): 
3095-3129. 

[7] Noroz Khan and Saleem Ahmed, "Finger-vein 
Image Dual Contrast Enhancement and Edge 

Detection," International Journal of Computer 
Science and Network Security 19, no. 11 
(2019): 184-192. 

[8] X. Qiu, W. Kang, S. Tian, W. Jia, and Z. 
Huang, "Finger Vein Presentation Attack 
Detection Using Total Variation 
Decomposition," in IEEE Transactions on 
Information Forensics and Security, vol. 13, 
no. 2, pp. 465-477, Feb. 2018.  

[9] B. T. Ton and R. N. J. Veldhuis, "A high-
quality finger vascular pattern dataset 
collected using a custom-designed capturing 
device," International Conference on 
Biometrics (ICB), pp. 1-5, Madrid, 2013.  

[10] Y. Lu, S. J. Xie, S. Yoon, Z. Wang and D. S. 
Park, "An available database for the research 
of finger vein recognition," 6th International 
Congress on Image and Signal Processing 
(CISP), pp. 410-415, Hangzhou, 2013.  

[11] W. Yang, X. Huang, F. Zhou, Q. Liao, 
"Comparative competitive coding for personal 
identification by using finger vein and finger 
dorsal texture fusion", Inf. Sci., vol. 268, pp. 
20-32, Jun. 2014. 

[12] Lee, Eui Chul, Hyunwoo Jung, and Daeyeoul 
Kim. "New finger biometric method using 
near infrared imaging." Sensors 11, no. 3 
(2011): 2319-2333. 

[13] Kejun, Wang, Liu Jingyu, P. Popoola 
Oluwatoyin, and Feng Weixing, "Finger vein 
identification based on 2-D gabor filter," In 
the 2nd International Conference on Industrial 
Mechatronics and Automation, vol. 2, pp. 10-
13. IEEE, 2010. 

[14] Ezhilmaran, D., and P. Rose Bindu Joseph, 
"Fuzzy inference system for finger vein 
biometric images." In 2017 International 
Conference on Inventive Systems and Control 
(ICISC), pp. 1-4. IEEE, 2017. 

[15] Lin, Cheng-Jian, Shiou-Yun Jeng, and Mei-
Kuei Chen, "Using 2D CNN with Taguchi 
parametric optimization for lung cancer 
recognition from CT images," Applied 
Sciences 10, no. 7 (2020): 2591. 

[16] Garg, Priyanka, and Trisha Jain, "A 
comparative study on histogram equalization 
and cumulative histogram equalization," 
International Journal of New Technology and 
Research 3, no. 9 (2017): 263242. 

[17] Shin, Kwang Yong, Young Ho Park, Dat Tien 
Nguyen, and Kang Ryoung Park, "Finger-vein 
image enhancement using a fuzzy-based 
fusion method with gabor and retinex 
filtering," Sensors 14, no. 2 (2014): 3095-
3129. 

[18] Pham, Tuyen Danh, Young Ho Park, Dat Tien 
Nguyen, Seung Yong Kwon, and Kang 
Ryoung Park, "Nonintrusive finger-vein 
recognition system using NIR image sensor 



Finger-Vein Image Dual Contrast Adjustment and Recognition Using 2D-CNN (pp. 16 - 25) 

Sukkur IBA Journal of Computing and Mathematical Science - SJCMS | Vol. 6 No. 1 January – June 2022 

25 

and accuracy analyses according to various 
factors," Sensors 15, no. 7 (2015): 16866-
16894. 

[19] Peng, Jialiang, Ning Wang, Ahmed A. Abd 
El-Latif, Qiong Li, and Xiamu Niu, "Finger-
vein verification using Gabor filter and SIFT 
feature matching," In 2012 Eighth 
International Conference on Intelligent 
Information Hiding and Multimedia Signal 
Processing, pp. 45-48. IEEE, 2012. 

[20] Pang, Shaohua, Yilong Yin, Gongping Yang, 
and Yanan Li, "Rotation invariant finger vein 
recognition," In Chinese Conference on 
Biometric Recognition, pp. 151-156. 
Springer, Berlin, Heidelberg, 2012. 

[21] Wu, Jian-Da, and Chiung-Tsiung Liu, 
"Finger-vein pattern identification using 
principal component analysis and the neural 
network technique," Expert Systems with 
Applications 38, no. 5 (2011): 5423-5427. 

[22] Xin, Yang, Zhi Liu, Haixia Zhang, and Hong 
Zhang, "Finger vein verification system based 
on sparse representation," Applied optics 51, 
no. 25 (2012): 6252-6258. 

[23] Syarif, Munalih Ahmad, Thian Song Ong, 
Andrew BJ Teoh, and Connie Tee, "Enhanced 
maximum curvature descriptors for finger 
vein verification," Multimedia Tools and 
Applications 76, no. 5 (2017): 6859-6887. 

[24] Radzi, Syafeeza Ahmad, Mohamed Khalil 
Hani, and Rabia Bakhteri, "Finger-vein 
biometric identification using convolutional 
neural network," Turkish Journal of Electrical 
Engineering & Computer Sciences 24, no. 3 
(2016): 1863-1878. 

[25] Gopinath, P., and R. Shivakumar. 
"Exploration of finger vein recognition 
systems." (2021). 

[26] Wang, Li, Haigang Zhang, and Jingfeng Yang, 
"Finger Multimodal Features Fusion and 
Recognition Based on CNN," In 2019 IEEE 
Symposium Series on Computational 
Intelligence (SSCI), pp. 3183-3188. IEEE, 
2019. 

[27] Ganesan, Thenmozhi, Anandha Jothi 
Rajendran, and Palanisamy Vellaiyan, "An 
Efficient Finger Vein Image Enhancement and 
Pattern Extraction Using CLAHE and 
Repeated Line Tracking Algorithm," In 
International Conference on Intelligent 
Computing, Information and Control Systems, 
pp. 690-700. Springer, Cham, 2019. 

[28] Sledevic, Tomyslav, "Adaptation of 
convolution and batch normalization layer for 
CNN implementation on FPGA," In 2019 
Open Conference of Electrical, Electronic and 
Information Sciences (eStream), pp. 1-4. 
IEEE, 2019. 

[29] Liu, Kui, Guixia Kang, Ningbo Zhang, and 
Beibei Hou, "Breast cancer classification 
based on fully-connected layer first 
convolutional neural networks," IEEE Access 
6 (2018): 23722-23732. 

[30] Borra, Surya Prasada Rao, N. V. S. S. Pradeep, 
N. Raju, S. Vineel, and V. Karteek, "Face 
recognition based on convolutional neural 
network," International Journal of 
Engineering and Advanced Technology 9, no. 
4 (2020): 156-162. 

[31] Noroz, Noroz Khan Baloch, Saleem Ahmed 
Saleem, and Ramesh Kumar Kumar. "Finger-
vein Image Enhancement and 2D CNN 
Recognition." International Journal of 
Innovations in Science & Technology 3 
(2021): 33-44