Intelligent human-centred mobile authentication system based on palmprints

Abstract

Biometrics, as an intelligent and secure authentication method, has recently become increasingly popular. Modern society uses fingerprints, iris and face recognition on a daily basis, even on a large scale; for example, in biometric passports. However, there are still other biometric traits that may provide sufficiently high accuracy but have not been widely implemented so far, e.g. palmprints. In this article, we propose a novel human-centred method of palmprint-based user verification. The proposed method is dedicated to the mobile devices and provides the accuracy reaching 94.5%. Moreover, the method is time-computing efficient and gives the response in less than 0.2 s. All the experiments described in the article were performed using the benchmark PolyU database and three widely available mobile phones.

Keywords

Mobile biometrics palmprint pattern recognition intelligent user authentication security image processing

1 Introduction

Biometrics has become increasingly popular nowadays, and now it is widely implemented in our daily life. Biometric solutions make human life easier; they make the verification more convenient and faster. What is more, biometrics increases the security of the stored data. Fingerprint is the most widely developed biometric trait and it has been used for person authentication for more than 100 years [11]. In spite of the popularity of fingerprints, some other traits have recently become more widely used: iris, gait and palmprint.

Palmprints meet all the requirements that have to be fulfilled by a potential biometric trait: 1) universality – each person has palmprints, 2) uniqueness – palmprints are different even in case of twins, 3) permanence – the palmprint pattern remains unchanged during a lifetime, 4) collectability – palmprints are easily collected and they may be obtained in a contactless way, 5) performance – the accuracy of the state-of-the-art methods is satisfying, 6) acceptability – palmprints do not refer to any intimate parts of humans body. The further advances are: stable and rich line features, small distortion and easy self-positioning [13]. All these above mentioned facts ensure that a palmprint is very promising for user authentication.

As presented in [14] and [22], biometrics may also be used as a part of the multi-factor authentication system. This approach may lead to the creation of the most secured system possible, because it considerably increases the spoofing effort for an attacker. However, in some cases the biometric artefact used in such a system may be attacked. Loss of biometric identity threatens a person’s security: it may cause social or financial losses and invades privacy, causing negative emotional impact on victims. Protecting biometric identities is particularly important in the today’s digital world because of the limited number of biometrics. Thus, some techniques for protecting the template are implemented, as presented in [3].

Another potential approach to template protection is called cancellable biometrics. When applied, the biometric sample becomes distorted or transformed in such a manner, that it becomes difficult to obtain the original template from the distorted one [18].

As mobile devices have become widely used, the biometric solutions moved to the mobile scenario. Nevertheless, using mobile devices poses different challenges. The crucial one is the still limited raw computing power available in these devices. It forces biometric application designers and developers to make some kind of compromise in terms of selecting and adapting the right algorithm to these platforms and the related operating systems [1].

In this article, we propose a novel, innovative and accurate approach to creating an intelligent and secure way of authentication. It uses compact code for biometric template representation. The proposed solution may be treated as the cancelable biometrics due to two facts. First of all, in the memory we store the code instead of the whole template. Secondly, the template cannot be rebuilt using the code. The proposed solution is dedicated to mobile devices and has been tested on several Android devices. This article is structured as follows: Section 2 refers to the state-of-the-art methods, Section 3 presents the overview of our proposed novel method, whereas Section 4 focuses on the obtained results. We provide some conclusions with the plan for the future work afterwards.

2 Related work

Recently, palmprint recognition has received some increasing research attention, and a variety of methods have been proposed for each part of the biometric authentication system: sample pre-processing, feature extraction and classification. Due to the very rich structure of a palmprint, there are multiple features that may be informative for the identification process: ridges, singular points, minutiae points, principal lines, wrinkles, palm texture, mean, variance, moments, center of gravity and density, spatial dispersivity and L1-norm energy, as reported in [28].

Pre-processing is essential for all the consecutive steps of the pipeline: feature extraction and classification. Its quality has a significant impact on the accuracy of recognition, as described in [32]. In this step the filtering, blurring and edges enhancement are mostly used. Nonetheless, pre-processing has one more important meaning. During this step the size of image is adjusted, and the ROI is selected, so that the computing is easier and faster. The example of ROI extraction algorithm was proposed in [9].

The most common methods for feature extraction are based on texture analysis. Among them Gabor filters used in [29] and developed in [25] can be listed. Moreover, researchers often propose using Fourier transform [7], Hough transform [2], scale invariant feature transform (SIFT) [12] or wavelets [39]. Texture analysis can perfectly represent the multiscale information of a palmprint image in frequency domain. After the transform, statistical indicators (i.e. variance, standard deviation, mean value, invariant moments, density) are computed and converted to a vector for matching.

Another approach to the feature extraction is the use of encoding-based methods. Among others, the following solutions may be enumerated: Palm Code [31], CompetitiveCode [37], Ordinal Code [26] and Binary Orientation Co-occurrence Vector [33]. The main idea of these algorithms is to reduce the complexity of image to the code (binary, numeric or alphanumeric). The code is easier and more time-efficient to deal with, especially when it comes to the processing on the mobile devices.

It is also possible to use histograms in order to get some information about palmprint ROI (Region of Interest). Histogram equalization is presented in [24], [8] or [34].

Autoencoders have gained raising interest recently. Commonly, they are a special type of artificial neural network used to learn efficient data coding in an unsupervised manner. As presented in [23] they can be successfully implemented for palmprint-based recognition.

When it comes to the classification step, there are mainly three possibilities: artificial intelligence techniques, machine learning classifiers or less complex measures. As an example of the first group may be listed a convolutional neural network used in [10] or [5]. Among classifiers the SVM method is often implemented like in. [36]. Nonetheless, there are still multiple approaches using less complex measures like the Euclidean distance in [38], Hamming distance in [17] or angular distance implemented in [20].

Recently, palmprint authorization has been moved to the mobile scenario more often. In [4], the pre-processing part of the method is described in detail. This step is designed in such a way, that it is able to perform segmentation of the hand and a complex background. In the feature extraction part, the Band-Limited Phase-Only Correlation (BLPOC) is employed. The ROI image is divided into nine 32×32px parts and BLPOC is calculated for each possible pair. In the classification part, the algorithm takes the highest peak value of the average BLPOC function.

The authors in [21] used the Orthogonal Line Ordinal Features (OLOF). The Authors of the OLOF method claimed an increased accuracy and reduced complexity in comparison with other feature extraction methods known from the literature. Thus, it seems to be a suitable method for the mobile scenario having limited computing power. As a result of the feature extraction step, three bit ordinal codes are generated. Then, in the matching step, they are compared to the pattern samples using the Hamming distance.

Another approach was presented by Kim et al.in [16]. There, the feature extraction step begins from a Gabor filter. In fact, there are six different orientations of filters and their real responses are further analyzed. The orientation that gives the maximum filter response is selected as the orientation of the pixel according to the selected competitive rule. Then, a palmprint image is divided into smaller blocks. For each of them, an orientation histogram is obtained by counting the number of pixels at each orientation. Next, histograms are used in order to create the feature vector. In this approach, the chi-square distance is used for classification.

In Tiwiari’s work [27], two well-known methods of feature extraction were used and merged together: Scale Invariant Feature Transform (SIFT) and Oriented FAST and Rotated BRIEF (ORB). As a result, both methods provide the interest point. In the classification step, the vector from the sample is compared to the vector obtained from the pattern, a genuine image. The threshold value was set up experimentally.

In the mobile approach it is also possible to implement the multimodal biometrics as presented in [6]. Authors introduced the fusion of palmprint and knuckle biometric in order to provided more efficient and more secure way of user authentication.

The last but not least important work is the one provided by Zhang et al. [35]. In this approach, the machine learning is implemented twice. Firstly, it is used for ROI extraction. The detector D was trained in order to detect the valley points of the humans hand. Secondly, the SiameseMobileNet is used for classification. The proposed method differs from the traditional approaches in a significant way. The Authors did not feed the neural network with a feature vector but with two images: a genuine sample and a testing sample.

The presented state-of-the-art review gives the general picture of the current situation regarding the palmprint-based verification systems. In Table 1, the methods used in feature extraction and classification steps are mentioned. Researchers work more on texture analysis than on any other kind of features (color, geometry). It is also very popular to merge texture analysis with the code creation, which can improve the security of the whole system. On the other hand, the mobile solutions are focused on finding the compromise between the accuracy and the computation time.

Table 1
The state-of-the-art verification systems

Authors Platform Feature extraction Classification

Aoyama et al. [4] mobile Band-Limited Phase-Only Correlation (BLPOC) Maximum of BLPOC

El-Tarhouni et. al. [7] PC Gabor filters and Local Binary Pattern KNN

Jaswal et. al. [12] PC SIFT and Texture Code Co-occurrence Matrix SVM

Moco et al. [21] mobile Orthogonal Line Ordinal Features (OLOF) Hamming distance

Kim et al. [16] mobile Bank of Gabor filters Chi-square distance

Tiwari et al. [27] mobile SIFT and ORB Own matching score

Zhang et al. [31] PC Gabor filters Hamming distance

Zhao et al. [37] PC Competitive Code Own matching score

Zhou et al. [39] PC Multi-wavelet Neural network

Authors	Platform	Feature extraction	Classification
Aoyama et al. [4]	mobile	Band-Limited Phase-Only Correlation (BLPOC)	Maximum of BLPOC
El-Tarhouni et. al. [7]	PC	Gabor filters and Local Binary Pattern	KNN
Jaswal et. al. [12]	PC	SIFT and Texture Code Co-occurrence Matrix	SVM
Moco et al. [21]	mobile	Orthogonal Line Ordinal Features (OLOF)	Hamming distance
Kim et al. [16]	mobile	Bank of Gabor filters	Chi-square distance
Tiwari et al. [27]	mobile	SIFT and ORB	Own matching score
Zhang et al. [31]	PC	Gabor filters	Hamming distance
Zhao et al. [37]	PC	Competitive Code	Own matching score
Zhou et al. [39]	PC	Multi-wavelet	Neural network

3 The proposed method for mobile palmprint-based authentication

In [19], an interesting approach to texture analysis was presented. The author introduced five different vectors. Using a pair of vectors, the two-dimensional kernel is created. Convolution of kernel and the source image may emphasize some specific characteristic (like spots, ripple or edges). In our research, we decided to use the vectors called L5 (Equation (1)) and S5 (Equation (2)). This pair of vectors gives one of two kernels (depending on the vector order – Equations (3) and (4)). In order to make the result rotational invariant, it is recommended to combine the symmetric pair of kernels. $L 5 = [1, 4, 6, 4, 1]$ (1) $S 5 = [- 1, 0, 2, 0, - 1]$ (2) ${kernel}_{LS} = [\begin{matrix} - 1, & 0, & 2, & 0, & - 1 \\ - 4, & 0, & 8, & 0, & - 4 \\ - 6, & 0, & 12, & 0, & - 6 \\ - 4, & 0, & 8, & 0, & - 4 \\ - 1, & 0, & 2, & 0, & - 1 \end{matrix}]$ (3)

In our proposed method, we used the convolution operations. Firstly, the ROI need to be extracted in the pre-processing part of algorithm. In order to perform the extraction, we run the algorithm presented in our previous work, which is described in detail in [30]. ${kernel}_{SL} = [\begin{matrix} - 1, & - 4, & - 6, & - 4, & - 1 \\ 0, & 0, & 0, & 0, & 0 \\ 2, & 8, & 12, & 8, & 2 \\ 0, & 0, & 0, & 0, & 0 \\ - 1, & - 4, & - 6, & - 4, & - 1 \end{matrix}]$ (4)

The most crucial steps of this algorithm are following. Firstly, thresholding is performed in order to get the foreground (hand) and the background. Then, the contour around the hand is extracted. Using the contour it is possible to detect two points in valleys between pairs of fingers: index and middle, and ring and little. The line is constructed between these two points and the region of interest is delimited in a specific offset from the line. After the ROI extraction, the convolution operations are performed. This part of the proposed method is presented in Fig. 1. The source image comes from the PolyU database – the ROI from the sample is presented in part A. Then, the result from the kernel_SL convolution is visualized in part B and the result kernel_LS in part C. Then, part D is the sum of B and C. As presented in the figure, the convolution operations emphasize significantly horizontal and vertical lines in the source image.

Fig. 1

Sequential steps of pre-processing: A – source image, B – LS filtered image, C – SL filtered image, D – sum of LS images B and C.

After applying kernel on the image, the texture energy measure (TEM) is calculated. It is computed by summing the absolute values in a local neighbourhood using Equation 5, where C(i; j) is the illumination of the pixel and i, j – x and y axis coordinates. During the research, we used the block sizes m = 15×n=15. It means that TEM is calculated 64 times for the source image, since the image size is 120px×120px. After the TEM measure is known for each block, the average value of TEM is calculated (TEMavg). Then, the 64-bit length binary vector of features is created using Equation (6). $TEM = \sum_{i = 1}^{m} \sum_{j = 1}^{n} | C (i . j) |$ (5) $vec [a] = {\begin{matrix} 0 & if TEM < TEM avg \\ 1 & otherwise \end{matrix}$ (6)

After implementing this kind of approach, we decided to run some experiments. The experiments are described in detail in the next section. Nevertheless, the obtained results were not satisfying. In order to improve the accuracy of the proposed system, we decided to add more features to the vector. As the convolution operation using kernels kernel_SL and kernel_LS emphasize the horizontal and vertical lines, we decided to extend the feature vector by adding those obtained from runlengths. Runlength is the feature that can provide multiple pieces of information about the analysed texture. Basically, in a course texture it is expected that the long runs will occur relatively often, whereas a fine texture will contain a higher proportion of short runs. Runlengths are analysed both horizontally and vertically, and they are the amount of consecutive primitives (pixels) having the intensity equal to 255 (white pixels in thresholded image). Six features were added to the vector: 1) number of runs having length _ 3, 2) number of runlengths > 60 and 3) the maximum runlength of image – analysed both horizontally and vertically. The whole system’s overview is presented in Fig. 2.

Fig. 2

The overview of the proposed method: the exemplary sample form PolyU database, extracted ROI, pre-processing step with convolution operations and thresholding, two types of code (short – obtained from convolution operations and long – the short code extended by runlengths), sample database (learning samples set), SVM classifier and the result.

Due to the limited computation power provided by the mobile devices, we decided to move one part of the system to the PC, as presented in Fig. 3. Thus, the training is performed using the computer and as a result an XML file is created.

Fig. 3

PC and mobile parts of the proposed system

Then, the file is moved to the mobile phone and the experiments are performed. For each testing sample, the system gives the answer: genuine or impostor.

4 Experimental setup and results

For evaluating our methods, we chose to use the PolyU palmprint database (http://www4.comp.polyu.edu.hk/ biometrics/), which was proposed in [31].

Currently, it is supposedly the most popular database for this biometric trait. For scientific purposes, it is available online after registration and ensures that the obtained results are comparable to the state-of-art approaches. The PolyU dataset contains the images of palmprints captured by a contact-based capture device. During the research, 7720 samples of palmprints were used: 20 samples for each of 386 users. We used the PC (Intel Core i5-4210U 1.7 GHz CPU, 4-Core, 4GB RAM, Windows 8.1-64 bit) for training and three different mobile phones for testing. The parameters of mobile devices are presented in Table 2.

Table 2
Specification of mobile devices used for the experiments and validation

No. Details

1. Samsung Galaxy A5 2017Android 8.0 Oreo8×1.90GHz3GB RAM

2. Xiaomi Mi6Android 7.1 Nougat8×2.45GHz6GB RAM

3. Samsung Galaxy S5Android 6.0 Marshmallow4×2.50GHz2GB RAM

No.	Details
1.	Samsung Galaxy A5 2017Android 8.0 Oreo8×1.90GHz3GB RAM
2.	Xiaomi Mi6Android 7.1 Nougat8×2.45GHz6GB RAM
3.	Samsung Galaxy S5Android 6.0 Marshmallow4×2.50GHz2GB RAM

For the first experiment, we used the 64-length vector of features (coming only from the TEM and convolution operations). In the experiments, we used a 10-fold classification – we run the experiment 10 times using a different set of training sets in order to provide

The lack of dependency on data. We decided to use 10 positive and 10 negative samples in the training step. The average accuracy (understood as an number of well classified samples divided by the total number of samples) of the method is 84.4%, so we did not find it satisfying. Therefore, we extended the vector and ran the experiment again. The average accuracy reached 94.5%, which gave over 10% of accuracy increase comparing to the shortest 64-bit vector. The more detailed results of both experiments are presented in Table 3.

Table 3

Obtained accuracy results for both feature vectors

Fold	Short code (64 elements)	Long code (70 elements)
#1	78.4%	95.6%
#2	84.4%	94.1%
#3	88.2%	96.6%
#4	88.1%	93.0%
#5	85.5%	94.9%
#6	84.6%	93.4%
#7	84.1%	96.6%
#8	81.3%	90.6%
#9	81.6%	96.1%
#10	87.7%	95.5%
Avg.	84.4%	94.5%

Apart from the accuracy, the operation time was assessed. Table 4 presents the time of processing one single authentication. For each fold, the average pre-processing time and the average feature extraction with classification time were calculated. The average time of the whole verification process (pre-processing+feature extraction+classification) was 181.99 ms for Samsung Galaxy A5 2017, 77.42 ms for Xiaomi Mi6 and 174.43 ms for Samsung Galaxy S5, respectively.

Table 4

Time of single verification on the mobile devices [ms]

	Samsung A5 2017		Xiaomi Mi6		Samsung S5
Fold	Pre-processing	Feature extraction+ classification	Pre-processing	Feature extraction+classification	Pre-processing	Feature extraction+ classification
#1	10.95	168.75	10.15	65.67	11.00	159.83
#2	12.89	172.42	9.54	65.08	15.07	168.92
#3	11.84	168.17	10.92	65.75	13.27	164.92
#4	11.21	167.42	11.08	66.33	13.07	154.08
#5	11.11	168.00	12.08	66.00	12.60	166.83
#6	10.11	168.50	11.15	66.67	11.53	151.33
#7	10.84	173.50	10.00	66.83	14.20	179.50
#8	10.79	173.67	13.54	66.25	13.87	144.83
#9	11.05	173.50	12.85	66.50	13.93	160.50
#10	11.84	173.33	13.38	65.42	11.80	163.25
Avg.	11.26	170.73	11.47	65.95	13.03	161.40

5 Conclusions and future work

In the article, we proposed a novel mobile approach to palmprint-based user authentication. It is a novel, innovative and accurate approach to creating an intelligent and secure way of authentication. The most important contribution of this research is a time-computing efficient method, reaching high accuracy – we report the accuracy equal to 94.5% and the response in less than 0.2 s for each of three testing devices. The method has been tested on the set of widely used smartphones (Samsung Galaxy A5, Xiaomi Mi6 and Samsung Galaxy S5). The proposed method, apart from the accuracy, also provides a high security level. It is important that only the classifier is stored on the mobile device. Having only the XML file with the classifier, the impostor cannot rebuild the whole biometric feature (the image of the palm).

When focusing on the accuracy, the proposed solution is comparable to the previously presented methods available in the literature.

In the Table 5, the details of the research are presented: database involved, accuracy or EER and the time of response (where provided). However, the proposed method outperforms the rest in the computing-time domain.

Table 5

The results of the state-of-the-art mobile palmprint-based verification methods and the proposed one

Authors	Accuracy or EER	Time of response	Database
Aoyama et al. [4]	EER=4.07%	1s	Own dataset
Moco et al. [21]	FRR=9.87%, FAR = 0.03%	466ms	Own dataset
Kim et al. [16]	EER=2.88%	685ms	Own dataset
Tiwari et al. [27]	Acc=94.64%	over 16s	Own dataset
Zhang et al. [35]	Acc=99.5%	not specified	Own dataset
Giełczyk at al. (proposed)	Acc=94.5%	less than 200ms	PolyU

We are going to continue our work in the palmprint-based identification field. This is reasonable due to the advantages of palmprint as the biometric trait. As presented in the article, it is possible to provide high accuracy and very short computing time, but there are some other crucial issues to concern, e.g. lack of proper, widely known and available for public database, that can be used in mobile scenario. Therefore, we plan to publish a new database, containing the images taken by mobile devices and dedicated strictly to the mobile, unsupervised scenario.

References

Abate

A.F.

, et al., I-Am: Implicitly authenticate me—Person authentication on mobile devices through ear shape and arm gesture, IEEE Transactions on Systems, Man, and Cybernetics: Systems 49(3) (IEEE 2017), 469–481.

Ahmadi

and Soleimani

, Palmprint image registration using convolutional neural networks and Hough transform, arXivpreprint arXiv:1904.00579, 2019.

Ali

S.S.

, et al., Securing biometric user template using modified minutiae attributes, Pattern Recognition Letters 129 (Elsevier 2020), 263–270.

Aoyama

, et al., A contactless palmprint recognition algorithm for mobile phones, International Workshop on Advanced ImageTechnology, 2013, pp. 409–413.

Chai

, et al., Boosting palmprint identification with gender information using DeepNet, Future Generation Computer Systems 99 (Elsevier 2019), 41–53.

Choraś

and Kozik

, Contactless palmprint and knuckle biometrics for mobile devices, Pattern Analysis and Applications 15(1) (2012), 73–85.

El-Tarhouni

, et al., Multispectral palmprint recognition based on local binary pattern histogramFourier features and Gaborfilter, 6th European Workshop on Visual Information Processing (EUVIP), IEEE 2016, pp. 1–6.

Giełczyk

, et al., Hybrid feature extraction for palmprint-based user authentication. International Conference on High Performance Computing & Simulation (HPCS), IEEE 2018, pp. 629–633.

Giełczyk

, et al., Towards Mobile Palmprint Biometric System with the New Palmprint Database. International Conference on Image Processing and Communications, Springer 2019, pp.149–157.

10.

Gong

, et al., Palmprint Recognition Based on Convolutional Neural Network-Alexnet, 2019 Federated Conference on Computer Science and Information Systems (FedCSIS), IEEE 2019, pp. 313–316.

11.

Ito

and Aoki

, Recent advances in biometric recognition, ITE Transactions on Media Technology and Applications 6(1) (2018), 64–80.

12.

Jaswal

, et al., Multiple feature fusion for unconstrained palmprint authentication, Computers & Electrical Engineering 72 (Elsevier 2018), 53–78.

13.

Jia

, et al., Palmprint recognition based on complete direction representation, IEEE Transactions on Image Processing 26(9) (IEEE 2017), 4483–4498.

14.

Kang

, et al., Safe and convenient personal authentication method using Moiré 3D authentication based on biometric authentication, Cluster Computing 22(11) (Springer 2019), 2017–2026.

15.

Kaur

and Khanna

, Random distance method for generating unimodal and multimodal cancelable biometric features, IEEE Transactions on Information Forensics and Security 14(3) (IEEE 2018), 709–719.

16.

Kim

J.S.

, et al., An empirical study of palmprint recognition for mobile phones, Transactions on Consumer Electronics 61(3) (IEEE 2015), 311–319.

17.

Kumar

, Toward more accurate matching of contactless palmprint images under less constrained environments. Transactions on Information Forensics and Security 14(1) (IEEE 2018), 34–47.

18.

Kumar

, et al., Cancelable Biometrics: a comprehensive survey, Artificial Intelligence Review, Springer 2019, pp. 1–44.

19.

Laws

K.I.

, Rapid texture identification, , Image processing formissile guidance 238 (1980), 376–381.

20.

, et al., Multi-orientation and multi-scale features discriminant learning for palmprint recognition, Neurocomputing 348 (Elsevier 2019), 169–178.

21.

Moco

N.F.

, et al., Smartphone-based palmprint recognition system, 21st International Conference on Telecommunications (ICT), IEEE 2014, pp. 457–461.

22.

Selwal

and Gupta

S.K.

, Low overhead octet indexed template security scheme for multi-modal biometric system, Journalof Intelligent & Fuzzy Systems 32(5) (IOS 2017), 3325–3337.

23.

Shao

, et al., Cross-domain palmprint recognition based on transfer convolutional autoencoder, International Conference on Image Processing (ICIP), IEEE 2019, pp. 1153–1157.

24.

Srivastava

, et al., Biometric authentication using local subspace adaptive histogram equalization, Journal of Intelligent & Fuzzy Systems 32(4) (IOS 2017), 2893–2899.

25.

Srivastava

, et al., Palmprint recognition using novel oriented Gabor gradients, Journal of Intelligent & Fuzzy Systems 35(5) (IOS 2018), 5345–5351.

26.

Sun

, et al., Ordinal feature selection for iris and palmprint recognition, IEEE Transactions on Image Processing 23(9) (IEEE 2014), 3922–3934.

27.

Tiwari

, et al., A palmprint based recognition system for smartphone, Future Technologies Conference (FTC), IEEE 2016, pp. 577–586.

28.

Unar

, et al., A review of biometric technology along with trends and prospects, Pattern Recognition 47 (Elsevier 2014), 2673–2688.

29.

Verma

and Chandran

, Contactless palmprint verification system using 2-D Gabor filter and principal component analysis, Int Arab J Inf Technol 16(1) (2019), 23–29.

30.

Wojciechowska

, et al., Evaluation of the pre-processing methods in image-based palmprint biometrics, International Conference on Image Processing and Communications, Springer 2017, pp. 43–48.

31.

Zhang

, et al., Evaluation of the pre-processing methods in image-based palmprint biometrics, International Conference on Image Processing and Communications, Springer 2017, pp. 43–48.

32.

Zhang

, et al., An optimized palmprint recognition approach based on image sharpness, Pattern Recognition Letters 85 (Elsevier 2017), 65–71.

33.

Zhang

, et al., Fragile bits in palmprint recognition, IEEE Signal Processing Letters 19(10) (IEEE 2012), 663–666.

34.

Zhang

, et al., Weighted multimodal biometric recognition algorithm based on Histogram of Contourlet Oriented Gradient feature description. Federated Conference on Computer Science and Information Systems (FedCSIS), IEEE 2019, pp. 381–384.

35.

Zhang

, et al., Pay By Showing Your Palm: A Study of Palmprint Verification on Mobile Platforms, International Conference on Multimedia and Expo (ICME), IEEE 2019, pp. 862–867.

36.

Zhong

, et al., Palmprint and dorsal hand vein dual modal biometrics. International Conference on Multimedia & Expo Workshops (ICMEW), IEEE 2018, pp. 1–6.

37.

Zhao

, et al., Palmprint recognition using a modified competitive code with distinctive extended neighbourhood, IET Computer Vision 12(8) (IET 2018), 1151–1162.

38.

Zheng

B.W.

, et al., Recognition method based on Gabor wavelet transform and discrete cosine transform, Engineering Letters 26(2), 2018.

39.

Zhou

, et al., Hierarchical palmprint feature extraction and recognition based on multi-wavelets and complex network, IET Image Processing 13(6) (IET 2018), pp. 985–992.

Intelligent human-centred mobile authentication system based on palmprints

Abstract

Keywords

1 Introduction

2 Related work

Table 2 Specification of mobile devices used for the experiments and validation No. Details 1. Samsung Galaxy A5 2017Android 8.0 Oreo8×1.90GHz3GB RAM 2. Xiaomi Mi6Android 7.1 Nougat8×2.45GHz6GB RAM 3. Samsung Galaxy S5Android 6.0 Marshmallow4×2.50GHz2GB RAM

References

Table 2
Specification of mobile devices used for the experiments and validation

No. Details

1. Samsung Galaxy A5 2017Android 8.0 Oreo8×1.90GHz3GB RAM

2. Xiaomi Mi6Android 7.1 Nougat8×2.45GHz6GB RAM

3. Samsung Galaxy S5Android 6.0 Marshmallow4×2.50GHz2GB RAM