Finger Type Classification with Deep Convolution Neural Networks

Yousif Ahmed Al-Wajih

, Waleed M. Hamanah

2,* a

, Mohammad A. Abido

2,3,4, b

Fouad AL-Sunni

and Fakhraddin Alwajih

Control and Instrumentation Engineering Department at KFUPM, Dhahran, 31261, Saudi Arabia

Interdisciplinary Research Center in Renewable Energy and Power Systems (IRC-REPS), KFUPM, Saudi Arabia

Department of Electrical Engineering at KFUPM, Dhahran, 31261, Saudi Arabia

K. A. CARE, Energy Research & Innovation Center (ERIC), KFUPM, Saudi Arabia

Faculty of Computers and Artificial Intelligence Cairo University, Giza, Egypt

Keywords: Artificial Intelligence, Deep Learning, Fingerprint Identification, Convolutional Neural Network.

Abstract: The Automated Fingerprint Identification System (AFIS) is a biometric identification methodology that uses

digital imaging technology to obtain, store, and analyse fingerprint information. There has been an increased

interest in fingerprint-based security systems with the rise in demand for collecting demographic data through

security applications. Reliable and highly secure, these systems are used to identify people using the unique

biometric information of fingerprints. In this work, a learning-based method of identifying fingerprints was

investigated. Using deep learning tools, the performance of the AFIS in terms of search time and speed of

matching between fingerprint databases was successfully enhanced. A convolutional neural network (CNN)

model was proposed and developed to classify fingerprints and predict fingerprint types. The proposed

classification system is a novel approach that classifies fingerprints based on figure type. Two public datasets

were used to train and evaluate the proposed CNN model. The proposed model achieved high validation

accuracy with both databases, with an overall accuracy in predicting fingerprint types at around 94%.

1 INTRODUCTION

Biometric information encompasses a set of unique

and measurable physical characteristics, including a

person’s fingerprints and particular facial features, as

well as one’s voice and handwriting. Each person’s

fingerprints are formed of unique shapes and curves

that remain unchanged during a person’s lifetime.

Hence, fingerprinting can quickly identify and

authenticate a person efficiently. Due to its evident

reliability in accurately identifying persons, biometric

information has become the focus of researchers and

companies specialized in protection technology.

Fingerprints are now being extensively used as a

simple means of authentication on smartphones and

other mobile devices.

A fingerprint is a biometric method utilized to

identify people and authenticate identities. Unique

features are extractable from the surface of each

https://orcid.org/0000-0002-5911-7364

https://orcid.org/0000-0001-5292-6938

Corresponding Author

fingerprint (Bian et al. 2019) and (Rani et al., 2019).

Many biometric techniques have been devised for

fingerprint recognition and identification using the

ridges and greyscale images. This work emphasized

using a deep learning algorithm and testing its ability

to perform this task. Fingerprints identification

methods have conventionally outperformed other

biometrics methods, such as face and speech

recognition, being well-established, reliable, and

robust (Minaee et al., 2019) and (Chaitra et al., 2021).

In this area, fingerprint orientation field estimation

typically improves the performance of automated

fingerprint identification systems.

Fingerprint identification systems typically encom-

pass fingerprint imaging, acquisition, preprocessing,

and feature extraction matching. A significant number

of studies focus on various aspects of fingerprinting

identification systems, including selection and

extraction of optimized features as well as different

proposed methods of matching (Valdes et al., 2019)

Al-Wajih, Y., Hamanah, W., Abido, M., Al-Sunni, F. and Alwajih, F.

Finger Type Classiﬁcation with Deep Convolution Neural Networks.

DOI: 10.5220/0011327100003271

In Proceedings of the 19th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2022), pages 247-254

ISBN: 978-989-758-585-2; ISSN: 2184-2809

247

and (Srivastava et al., 2022). The old technique was

utilized based on deep learning in order to distinguish

four classes (arch, tented arch, left loop, right loop, and

whorl), using the Galton-Henry classification in (Shea,

2009 and Srivastava et al., 2021).

Fingerprints and facial features are presently the

most thoroughly studied biometric indicators, allowing

for reliable recognition in various applications. There

has been a growing need for more accurate and reliable

biometric identification and authentication-based

models from smartphones to border control. Recently,

researchers have been able to enhance the robustness

of recognition and identification models by

incorporating deep learning (Ribeiro et al., 2018) and

(Ayan et al., 202). In the following segments, we

review a number of the most recent related studies.

The authors in (Stojanovic et al., 2017) reviewed

recent methods in identifying latent fingerprints and

compared the most recent minutia descriptors. They

reported that selecting a good minutia would result in

improved accuracy of the AFIS. Their work detailed

the various minutia descriptors that could be used in

automatic fingerprint feature extraction and

compared them in terms of identification rates. They

proposed that the minutia descriptor C&J - which

relies on deep learning algorithms - be the new focus

of research in the area of latent fingerprint

identification. Furthermore, the authors

recommended that new minutia descriptors based on

deep learning be developed and the identification

accuracy be studied. In particular, they recommended

conducting studies to identify the best minutia

descriptors to enhance the performance of AFISs.

In (Preetha and Sheela, 2018), the authors

suggested that understanding the advantages and

limitations of the fingerprint orientation field

estimation methods is of fundamental importance to

creating fingerprint identification. According to the

authors, a common misunderstanding is that

automatic fingerprint identification had not been

appropriately addressed, despite AFIS being a subject

of research for decades. They explained that

fingerprint identification remains a significant pattern

identification dilemma of interest to researchers due

to the large intra-class mutability and inter-class

relationships in fingerprint patterns. They stressed

that automatic fingerprint identification systems

typically attempt to ascertain reliable matching

features from fingerprint images of inferior quality or

latent images, ‘damaged’ and ‘defects’ such as scars,

dirt, grease, and/or moisture on the surface of

fingertips. In (Cao and Jain, 2015), the authors

concluded that learning-based methods based on deep

learning had significantly improved the performance

of fingerprint orientation field estimation systems,

especially when dealing with challenges that

traditional methods had typically failed to tackle,

including latent fingerprints (such as poor-quality

fingerprints). They summarized the limitations of

conventional techniques as follows: 1) the initial

orientation fields are typically unreliable; 2) relying

primarily on high-quality fingerprints, their

algorithms may fail to handle latent and poor-quality

fingerprints; 3) human intervention during the

process of algorithm execution may be required; 4)

high computational complexity of such approaches.

In (Cao and Jain, 2019), the authors studied using a

CNN in running the fingerprint estimation algorithm

by modeling orientation field estimations of a poor-

quality image patch as a classification mission. They

classified the latent patch as one of a set of illustrative

orientation patterns using a CNN. The CNN was able

to learn the input images' characteristic features

directly. The authors concluded that fingerprint

identification estimated through a CNN would result

in higher accuracy than dictionary-based methods.

Schuch et al. 2017 trained CNNs as regression

networks to assess a fingerprint orientation field.

They called this proposed model a ConvNetOF. The

most recent work done in fingerprint classification

using the DL method is reported in (Michelsanti et al.,

2017), (Peralta et al., 2018), and (Zia et al., 019). DL-

based methods have been recognized as powerful

tools in the classification field (Lecun et al0., 2015).

Despite the fact that the wide use of DL approaches

in image classification, there remains a research gap

with regards to their use in fingerprint classification.

In that regard, the early work on this field was started

by authors in (Shea, 2009), (Wang et al., 2016), and

(Kakadiaris et al. 2009).

The most recent works on fingerprint

classification with new deep learning techniques are

considered in (Michelsanti et al., 2017); two pre-

trained CNN models (VGG) were evaluated using the

National Institute of Standards and Technologies

(NIST) SD4 dataset. The proposed models were

compared in terms of fast feature extraction. The

authors showed that DL-based methods outperformed

other methods due to their learning ability from the

row data. In addition, a deep CNN (DCNN) was used,

and the reported accuracy stood between 88.9% and

90% with the same NIST SD4 dataset 0(Peralta et al.,

2018). Further in (Zia et al., 2019), the authors

proposed a baseline DCNN model, and the reported

accuracy stood between 92.2% and 96.1% with the

NIST SD4 dataset. In addition, the authors reported

the high robustness of the proposed model. In (Blanco

et al., 2020), basic and modified extreme learning

ICINCO 2022 - 19th International Conference on Informatics in Control, Automation and Robotics

248

machines (ELMs) were tested for their efficacies

concerning fingerprint classification. The authors

showed that the improved ELM had outperformed the

other CNN models in terms of training speed and

computational cost.

Furthermore, the authors reported that the enhanced

ELM was able to handle data with the unbalanced class

distribution. They said the accuracy of 95%. They

concluded that the weighted ELM had achieved better

results in terms of accuracy and penetration rate

metrics. In (Iloanusi and Ejiogu, 2020), the work

focused on classifying input fingerprint grayscale

images according to the gender of the person being

identified. The authors reported an overall accuracy

rate of 91.3% in the classification. In this study, a 20-

layer CNN model was used. The model used was built

from the ground up. They employed both a Sokoto

Coventry Fingerprint Dataset (SOCOFing) dataset and

their dataset for training and testing. All previous work

focused on typically studied categories of the old four

classes (arch, tented arch, left loop, right loop, and

whorl). The paper focuses on fingerprint type.

Therefore, in this brief, labeling the datasets and

utilizing the state-of-the-art deep learning technique

with CNN structure is conducted to classify a

fingerprint type. As presented in the literature review,

and to the best of our knowledge, no work has

previously tackled finger type classification, which

marks this work's novelty.

In this study, the proposed new classifier was

designed to identify fingerprints as either thumb or

non-thumb. Such classification will improve the

matching time and the accuracy of AFIS. The data

have been labeled in the two-class. Then, training and

validation of the data have been applied. Deep

learning was used to classify the gray image of

fingerprint. A model of CNN was applied using the

benchmarked dataset. The proposed DL model is

used to help the matching algorithm verify the input

fingerprint more expediently, as it would require the

matching algorithm to search on half of the database.

This paper is organized as follows: In Section 2,

the data preparation with the proposed structure is

involved. The CNN architecture model is presented

and discussed in Section 3. Then, Section 4 presents

the experimental results for Thumb CNN (TCNN)

model. Finally, a conclusion is derived in Section 6.

2 PROPOSED SCHEME AND

DATA PREPARATION

The proposed model in this work is used to classify

the fingerprint image based on the finger type, as

shown in Figure 1. First, the dataset has is prepared

and labeled into the target classes. The labeled dataset

is separated into training and validation sets. A part

of the dataset is reserved for measuring the

performance of the proposed model. Second, the deep

learning model based on CNN structure is

investigated. Then, the proposed model is trained and

tuned in order to achieve high classification accuracy.

Finally, the unseen dataset is used to test the model

performance.

Figure 1: CNN Proposed scheme.

Two benchmark datasets were used to evaluate

the performance of our proposed model. The first one

was the NIST dataset (NIST, Biometric Special

Databases and Software, 2022). This database

featured 4000 8-bit grayscale 512x512 pixel

fingerprint images at the time of the study. The

dataset was collected randomly and stored in PNG

format. This dataset has been widely used in testing

and developing automated fingerprint classification

systems. The original database was classified into

five categories (L = left loop, W whirl, R = right loop,

T = tented arch, and A = arch). Subsequently, the

dataset is reorganized to match the newly proposed

classes. The naming scheme of the PNG files was

done such that the two numbers after the underscore

indicate the finger type and from the hand from which

the fingerprint image was taken. For example, in the

file labeled ‘f0001_01.png,’ the ‘01’ after the

underscore indicates that the fingerprint belongs to a

right thumb (Karu and Jain, 1996)0. The dataset was

divided into five files, one for each finger type. For

the thumb classifier, the same dataset was utilized. In

this case, thumb data was used, and the non-thumb is

collected randomly from the index, middle, ring, and

little fingers. Thumb sample data included 700

samples, 600 of which were set for training and 100

for validation and testing. An equal number of 175

samples were randomly selected from each class

(index, middle, ring, and little fingers). Data in this

model was divided into 85% for training and 15% for

validation. Samples from the datasets are illustrated

Figure 2.

Finger Type Classiﬁcation with Deep Convolution Neural Networks

249

Figure 2: Samples from the Datasets.

The second dataset used to train and test our proposed

models was the SOCOFing dataset (Shehu et al.,

2019). The study consisted of actual fingerprints

taken from 600 subjects at the time of the study.

Images were labeled according to the exclusive

attributes of gender, hand, and finger type. The real

part of this dataset was used for the purposes of this

study. The dataset was divided into subclasses in

order to create both thumb and other finger-type

models. For the thumb model, the dataset was

clustered into two classes of a total of 1200 thumb

fingerprint images of the BMP format. The non-

thumb class consisted of 300 images from the index,

middle, ring, and little fingers, for 1200 images. Of

this dataset, 75% was used for training and 25% for

validation.

3 CNN ARCHITECTURE

As discussed in the previous sections, the study was

to classify fingerprint data into subclasses. Our

approach in this work was to use a learning-based

method with a supervised learning methodology. To

this end, a deep learning technique was utilized to

achieve the classification target. As a state-of-the-art

model of deep learning and machine learning, CNN

was determined as ideal for classification tasks of

image-based data (Shyu et al., 2020). The architecture

of the thumb CNN (TCNN) model used in this work

is described in the coming subsection.

A CNN model was developed to train the

classification model. The model consisted of four

convolutional layers, each of which was followed by

a max-pooling layer. Filters in the four convolutions

numbered 256, 128, 64, and 32, respectively. Type

3X3 filters and a ReLU activation were used. The

application of consecutive convolutional and max-

pooling layers resulted in tensors of size (6, 6, 32),

which were flattened to size (1,152). Two dense

layers 128 and 20 neurons in size were then added.

The fully connected layers were supplied with ReLU

and Softmax activations consecutively. We used a

dropout layer between the two fully connected layers

with a drop rate of 30%. For training, we used cross-

entropy as the loss and an Adam optimizer in (Shyu

et al., 2020) for the backpropagation algorithm. Keras

with TensorFlow backend was used to create and

train the CNN model. The model summary is shown

in Figure 3.

Figure 3: TCNN Model Summery.

ICINCO 2022 - 19th International Conference on Informatics in Control, Automation and Robotics

250

4 EXPERIMENTS AND RESULTS

The proposed technique has been tested based on the

preparation data, and all experiments were conducted

using Python with TensorFlow and Keras libraries.

Training the models was conducted using Google

Collaboratory GPU resources. The results of the

training and testing TCNN model will be detailed in

the subsequent sections.

4.1 Training TCNN Model

Input data of both datasets used for the purposes of

training and testing are summarized in Table 1. The

NIST D4 dataset was randomly distributed into 70%

for training, 15% for validation, and 15% for testing.

At the same time, SOCOFig datasets were randomly

distributed into 75% for training, 15% for verification,

and 10% for testing. Before training the model, data

augmentation was employed and tuned in order to

increase accuracy and prevent overfitting. Images

used in training rotated within 20

, shifted right and

left within 10%, with image shearing and zoom

within 10%, and with horizontal flipping—the use of

the aforementioned augmentation technique allowed

for the enhancement of all models. The best accuracy

and loss metrics results stood at a 97% validation

accuracy, a 0.13 validation loss with the NIST D4

dataset, a 96% validation accuracy, and a 0.1

validation loss with the SOCOFing dataset. Results

are detailed in

Table 1

; also, Figure 4 and Figure 5

illustrate the training performed on the mentioned

dataset.

Table 1: Summary of The Two Datasets.

Dataset Training Validation Test

NIST SD4

498 thumbs

498 not-thumb

• 125 indexes

• 125 middles

• 124 rings

• 124 little

102 thumbs

102 not-thumb

• 26 indexes

• 26 middles

• 25 rings

• 25 little

100 thumbs

100 not-thumb

• 25 indexes

• 25 middles

• 25 rings

• 25 little

SOCOFing

900 thumbs

900 not-thumb

• 225 indexes

• 225 middles

• 225 rings

• 125 little

180 thumbs

180 not-thumb

• 45 indexes

• 45 middles

• 45 rings

• 45 little

120 thumbs

120 not-thumb

• 30 indexes

• 30 middles

• 30 rings

• 30 little

(a)

(b)

Figure 4: Results of the TCNN Model based on NIST SD4:

(a) model loss and (b) model accuracy.

(a)

(b)

Figure 5: Results of the TCNN Model based on SOCOFing:

(a) model loss and (b) model accuracy.

Finger Type Classiﬁcation with Deep Convolution Neural Networks

251

Table 2: Training of TCNN Model.

Dataset

Training Validation

Accuracy Loss Accuracy Loss

NIST D4 89.43 0.23 86.90 0.21

SOCOFing 91.97 0.18 92.26 0.21

Table 3: Matric Results for Test Set.

Metric NIST D4 SOCOFing

Accuracy 90.00% 89.00%

Precision 95.28% 83.63%

Recall 84.16% 92.00%

F1-score 89.38% 87.61%

4.2 Testing CNN Model

Loss and accuracy results of both the training and

validation sets are illustrated in Figure 4 and Figure 5.

Accuracy results of the validation set are summarized

in Table 2. Regularization induced by the dropout

layer allowed for more extended training of the model

and reduced the possibility of overfitting. Table 3

shows the unseen test dataset's accuracy, precision,

recall, and F1 scores. Notably, accuracy is decent for

a classification problem. Other metrics indicated that

predictions were somewhat uniform across the

different classes. In order to check how our model had

performed concerning individual classes, we used

confusion matrices. Each matrix showed the correctly

classified samples in the diagonal, according to class;

it also gave an insight into what classes are confused

by the model. Our model performed superiorly in

terms of differentiation between classes, as shown in

the diagonal of the confusion matrix in Figure 6.

To the best of our knowledge, there is no work in

the literature tackling the classification of the

fingerprint image to the finger type (thumb or not

thumb). However, there are some works have been

done on the same dataset with different problems

which are not comparable with our proposed work.

Table 4 summarizes the best result achieved in the

literature of three different field and features. The best

accuracy achieved in classifying fingerprints to

Galton-Henry classification (arch, tented arch, left

loop, right loop, and whorl) is 95.05 % (Michelsanti

et al. 2017). The best accuracy reached in assigning

gender (Male or Female) from fingerprints image is

91.3% (Iloanusi and Ejiogu, 2020). The accuracy

achieved in classifying that fingerprint is for a right

hand or left hand is 96.80% (Kim et al., 2020) where

this accuracy is a validation accuracy during the

training process not a test accuracy on an unseen

dataset.

(a)

(b)

Figure 6: Confusion Matric (a) SOCOFing test set. (b)

NIST D4 test set.

Table 4: Tackled PROBLEMS IN Literature.

Tackled Problem

Accuracy

(%)

classifying fingerprints into arch,

tented arch, left loop, right loop, and whorl

(Michelsanti et al. 2017).

95.05

Gender classification (Iloanusi and

Ejiogu, 2020)

91.30

Left- or Right-Hand Classification

(Kim et al., 2020)

96.80

Finger Type Classification 90.00

5 CONCLUSION

A novel approach for classifying fingerprints based

on the finger type was introduced through this work.

Results of implementation and experimentation

indicated that the TCNN model performed superiorly,

with a high accuracy rate of fingerprint type

ICINCO 2022 - 19th International Conference on Informatics in Control, Automation and Robotics

252

classification. The deep learning technique evidently

aided in the proper extraction and classification of

fingerprints. The developed model was trained and

evaluated using two datasets, NIST and SOCOFing.

The main metrics considered in this work, commonly

considered in studies of DL/CNN architecture, were

chosen to best reflect the level of performance in

terms of classification and features extraction. The

proposed model was able to classify the type of the

fingerprint with the accuracies of 90% and 89% with

the NIST D4 and SOCOFing datasets, respectively.

ACKNOWLEDGEMENTS

The authors would like to acknowledge the support

provided by King Fahd University of Petroleum &

Minerals. The authors also acknowledge the support

by KACARE Energy Research & Innovation Center

(ERIC) at KFUPM.

REFERENCES

Bian W., Xu D., Li Q., Cheng Y., Jie B., and Ding X.,

(2019). A Survey of the methods on fingerprint

orientation field estimation, IEEE Access, vol. 7, pp.

32644–32663.

Rani S., Kumar P., (2019). Deep Learning Based Sentiment

Analysis Using Convolution Neural Network, Arab J

Sci Eng 44, 3305–3314. https://doi.org/10.1007/s133

69-018-3500-z

Minaee S., Abdolrashidi A., Su H., Bennamoun M., and

Zhang D., (2019). Biometric Recognition Using Deep

Learning: A Survey, no. December.

Chaitra Y. L., Dinesh R., Gopalakrishna M. T., Prakash B.

V., (2021). Deep-CNNTL: Text Localization from

Natural Scene Images Using Deep Convolution Neural

Network with Transfer Learning, Arab J Sci Eng.

https://doi.org/10.1007/s13369-021-06309-9.

Valdes-Ramirez D. et al., (2019). A Review of Fingerprint

Feature Representations and Their Applications for

Latent Fingerprint Identification: Trends and

Evaluation, IEEE Access, vol. 7, pp. 48484–48499.

Srivastava, Hari M., Khaled M. Saad, and Walid M.

Hamanah, (2022), Certain New Models of the Multi-

Space Fractal-Fractional Kuramoto-Sivashinsky and

Korteweg-de Vries Equations" Mathematics 10, no. 7:

1089. https://doi.org/10.3390/math10071089.

Shea J. J., (2009). Handbook of fingerprint recognition

[Book Review], vol. 20, no. 5.

Srivastava, Hari M., Abedel-Karrem N. Alomari, Khaled

M. Saad, and Waleed M. Hamanah, (2021), Some

Dynamical Models Involving Fractional-Order

Derivatives with the Mittag-Leffler Type Kernels and

Their Applications Based upon the Legendre Spectral

Collocation Method" Fractal and Fractional 5, no. 3:

131. https://doi.org/10.3390/fractalfract5030131.

Ribeiro J. Pinto, Cardoso J. S., and Lourenco A., (2018).

Evolution, current challenges, and future possibilities in

ECG Biometrics,” IEEE Access, vol. 6, pp. 34746–

34776.

Ayan E., Karabulut B., Ünver H. M., (2021). Diagnosis of

Pediatric Pneumonia with Ensemble of Deep

Convolutional Neural Networks in Chest X-Ray

Images, Arab J Sci Eng. https://doi.org/10.1007/

s13369-021-06127-z.

Stojanovic B., Marques O., and Neskovic A., (2017). Latent

overlapped fingerprint separation: a review, Multimed.

Tools Appl., vol. 76, no. 15, pp. 16263–16290.

Preetha S. and Sheela S. V., (2018). Selection and

extraction of optimized feature set from fingerprint

biometrics-a review, Proc. 2nd Int. Conf. Green

Comput. Internet Things, ICGCIoT 2018, pp. 500–503.

Cao K. and Jain A. K., (2015). Latent orientation field

estimation via convolutional neural network, Proc.

2015 Int. Conf. Biometrics, ICB 2015, pp. 349–356.

Cao K. and Jain A. K., (2019). Automated Latent

Fingerprint Recognition, IEEE Trans. Pattern Anal.

Mach. Intell., vol. 41, no. 4, pp. 788–800.

Schuch P., Schulz S. D., and Busch C., (2017) Convnet

regression for fingerprint orientations, Lect. Notes

Comput. Sci. (including Subser. Lect. Notes Artif.

Intell. Lect. Notes Bioinformatics), vol. 10269 LNCS,

pp. 325–336.

Michelsanti D., Ene A. D., Guichi Y., Stef R., Nasrollahi

K., and Moeslund T. B., 92107). Fast fingerprint

classification with deep neural networks, VISIGRAPP

2017 - Proc. 12th Int. Jt. Conf. Comput. Vision,

Imaging Comput. Graph. Theory Appl., vol. 5, no.

Visigrapp, pp. 202–209.

Peralta D., Triguero I., Garcia S., Saeys Y., Benitez J. M.,

and Herrera F., (2018). On the use of convolutional

neural networks for robust classification of multiple

fingerprint captures, Int. J. Intell. Syst., vol. 33, no. 1,

pp. 213–230.

Zia T., Ghafoor M., Tariq S. A., and Taj I. A., (2019).

Robust fingerprint classification with Bayesian

convolutional networks, IET Image Process., vol. 13,

no. 8, pp. 1280–1288.

Lecun Y., Bengio Y., and Hinton G., (2015). Deep learning,

Nature, vol. 521, no. 7553, pp. 436–444.

Wang R., Han C., and Guo T., (2016) A novel fingerprint

classification method based on deep learning, Proc. -

Int. Conf. Pattern Recognit., vol. 0, pp. 931–936.

Kakadiaris I. A., Passalis G., Toderici G., Perakis T., and

Theoharis T., (2009). Face Recognition, 3D-Based,

Encycl. Biometrics, pp. 329–338.

Blanco D., Mora M., Barrientos R. J., Hernandez-García R.,

and Naranjo-Torres J., (2020). Fingerprint

classification through standard and weighted extreme

learning machines, Appl. Sci., vol. 10, no. 12.

Iloanusi O. N. and Ejiogu U. C., (2020). Gender

classification from fused multi-fingerprint types, Inf.

Secur. J., vol. 29, no. 5, pp. 209–219.

Finger Type Classiﬁcation with Deep Convolution Neural Networks

253

NIST, (2022). Biometric Special Databases and Software |

NIST,” [Online]. Available: https://www.nist.gov/itl/

iad/image-group/resources/biometric-special-

databases-and-software.

Karu K. and Jain A. K., (1996). Fingerprint classification,

Pattern Recognit., vol. 29, no. 3, pp. 389–404.

Shehu Y. I., Ruiz-Garcia A., Palade V., and James A.,

(2019). Detailed Identification of Fingerprints Using

Convolutional Neural Networks, Proc. - 17th IEEE Int.

Conf. Mach. Learn. Appl. ICMLA 2018, pp. 1161–

1165.

Shyu M., Chen S., and Iyengar S. S., (2020). a Survey on

Deep Learning Techniques, Strad Res., vol. 7, no. 8.

Kim, Junseob, Beanbonyka Rim, Nak-Jun Sung, and Min

Hong. (2020). Left or right hand classification from

fingerprint images using a deep neural network, CMC-

Computers Materials & Continua, v.63, no.1, pp.17 –

30.

ICINCO 2022 - 19th International Conference on Informatics in Control, Automation and Robotics

254