An Accurate Approach to Classify Real Time Indian Twins Using

SVM and Compare the Performance over Logistic Regression

Vallipi Dasaratha and J. Joselin Jeya Sheela

Department of Electronics and Communication Engineering, Saveetha Institute of Medical and Technical Sciences,

Saveetha University, Chennai, Tamil Nadu, 602105, India

Keywords: Biometric, Face Recognition, Identification, Image Processing, Logistic Regression, Support Vector

Machine, Technology.

Abstract: The study aims to increase the accuracy of comparing and classifying the Real Time Indian Twins using

Support Vector Machine (SVM) over Logistic Regression (LR) algorithm. Face Recognition of twins with

face and ID recognition using Support Vector Machine(SVM) over Logistic Regression. Here the analysis

was carried out with two groups named as Group 1 and 2 with sample iteration of 40 where each group consist

of 20 sample iterations, for a sample size of 1430. Results and Discussion: Compare and identify The Real

Time Indian Twins and also its Performance using SVMand Logistic Regression Algorithms. The SVM and

LR have achieved the accuracy of 62.2650% and 31.0225%. respectively. By comparing the accuracy of the

two algorithms, independent samples tests reveal an accuracy gap between the two methods that is statistically

significant of p=0.001 (p<0.05) which shows that the hypothesis is significant and is carried out using an

independent sample T- test. Conclusion: The findings clearly demonstrates that SVM has an excellent

accuracy of 62.265% when compared to Logistic Regression whose accuracy is 31.0225%.

1 INTRODUCTION

To generate a biometric identification code, facial

recognition scans a face against a template that has

been saved. Both identification and verification can

be accomplished using face recognition. Biometrics

is the use of biological or psychological data. The

human traits of an individual might reveal their

identity. At the most essential places, such as banks,

airports, enterprises, and many others, identification

of a person is performed because it is quite

important(Reddy, Siva Malleswar Reddy, and

Poovizhi 2022). There are numerous ways to do the

task, including knowledge-based processing, which is

also known as assigning an identity and password.

However, these methods have drawbacks. Twin

identification is made possible by a variety of facial

recognition technologies that also include voice, iris,

and other types of recognition. To identify a unique

person, the finger-print identification technique is

used(Suguna, Raja Suguna, and Amaresh 2022). The

disadvantages of biometric fingerprint technology

recognition include vulnerability to hacking and

delayed processing. Additionally, there are

limitations to iris recognition, such as identification

errors brought on by mismatching, a drawn-out

identification procedure, and unsuitability for

identification if any have eye defects (Babaeian et al.

2019) (Mahapatra, S. et al., 2016). Voice recognition

has drawbacks including being readily abused by

others. The limitation for facial recognition when two

different people have a striking resemblance in look

is called "identical twins."(Ozdemir and Tugrul

2019). If one topic is attempting to pass as another,

the two subjects may look remarkably similar. It is

crucial to test facial recognition algorithms on the

most challenging situations. Therefore, the new

approach is suggested to accurately identify twins

(Lee and Kwon 2018).

An image of a face is used as the input for this job.

To extract characteristics from the photos, the logistic

regression method is utilized (Agarwal et al. 2020).

The SVM classifier is then used in order to classify.

This makes twin detection straightforward, and the

kNN classifier will eliminate trouble because of how

similar the facial photographs are. Recent research

have revealed that automatic facial recognition

technology performs noticeably worse when images

belong to identical twin brothers than to unrelated

individuals (Ramkumar, G. et al. 2021). The

378

Dasaratha, V. and Sheela, J.

An Accurate Approach to Classify Real Time Indian Twins Using SVM and Compare the Performance over Logistic Regression.

DOI: 10.5220/0012772700003739

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 378-384

ISBN: 978-989-758-661-3

degradation of the face is shown to be far worse than

that of other biometrics, such as fingerprint and iris

scanning.

Face recognition software for people Since

humans are very adept at distinguishing persons from

photographs, performance is often used as a

benchmark for face recognition systems .To the best

of knowledge, no extensive research on humans been

conducted to address the issue of separating identical

twins solely only on facial images (Ptaszynski and

Masui 2018) (Deena et al. 2022). Here, test people's

ability to tell apart identical twin siblings or the same

person when they see two facial images. The ability

of humans to discern between the facial photographs

of identical twin siblings suggests that they may be

able to spot distinguishing characteristics that might

be utilized to enhance the accuracy of face

recognition technology software. In this study, human

volunteers look at pairs of facial pictures and rate how

positive they are that the faces belong to the same

individual or are identical twins (Subasi 2020).

2 MATERIALS AND METHODS

The dataset for this particular research study was

gathered from the Twinsburg Festival in Ohio. 25%

of the database is set aside for testing, while 75% is

reserved for training. A total of forty samples are

taken into account, divided into two sets, each of

which has twenty data samples. The G power was set

at 80%, the confidence interval was set at 95%, and

the threshold for the computation was set at where

Group 1 was a Logistic Regression (LR) method and

Group 2 was SVM algorithm, which includes 20

iterations each with the total of 40 iterations for 1430

sample size.

2.1 SVM and Logistic Regression

An edge detection linear filter used in image

processing is called a Gabor filter. Gabor filters are

band pass filters used in image processing for the

computation of stereo disparity, feature extraction

from textures, and texture analysis. A complex

oscillation is multiplied by a Gaussian envelope

function, these filters create impulsive responses.

These fundamental functions, according to Gabor, the

space-time-uncertainty product should be reduced.

Orientation-selective filters can be produced by

extending the functions to two dimension. The

response phase of the Gabor filter can occasionally be

roughly linear. By comparing the phase difference of

the left and right filter response, The disparity in the

stereo images is estimated using stereo techniques

that make advantage of this attribute. According to a

number of studies, the profile of simple-cell receptive

field in the mammalian brain can be well represented

by two-dimensional Gabor function.

Algorithm Steps for Logistic Regression

1: Input: Training data

2: Begin

3: For i = 1 to k

4: For each training data instance di.

5: Set the target value for the regression to zi = Yi-

P(1d)/[P(1|d,)(1−P(1]dj))]

6: Initialize the weight of instance d; to [P(1d;)(1 −

P(1d;))]

7: Finalize a ƒ (j) to the data with class value (Zj) and

weight (wj)

8: Classical label decision

9: Assign (class label: 1) if Pid > 0.5, otherwise (class

label: 2)

10: End

The number of scales and orientations, as well as

the maximum and minimum centre frequency, is used

to construct the spatial domain of the Gabor wavelets.

Ventral stream from v1 to v2 to v4 IT.Units of V1

implement Gabor filters (in S1 layer of). This layer

heavily filters an input grayscale image (120x120 or

160x160 pixels) using a range of Gabour filters in

various scales and orientations. As a result, a filter of

every size and orientation is centred on each pixel of

the input image. The filters are supplied in 4

orientations with 8 bands and 16 scales (resulting in

64 maps when multiplied by 4).

2.2 Pseudocode of the Support Vector

Machine (SVM) Algorithm

Input: D=[X,Y];X(array of input with m features),

Y(array of class labels)

Y=array(C)// Class label

Output: Find the performance of the system

function train_svm(X,Y,number_of_runs)

initialize:learning_rate=Math.random();

for learning_rate innumber_of_runs

error=0;

An Accurate Approach to Classify Real Time Indian Twins Using SVM and Compare the Performance over Logistic Regression

379

for i in X

if (Y[i]*X[i]*w))<1

The update: w=w+learning_rate*((X[i]*Y[i])*(-

2*(1/number_of_runs)*w)

Else

update:w=w+learning_rate*(-

2*(1/number_of_runs)*w)

end if

end

End

This describes the division of all classes into two

distinct categories. In the root-node of the decision

tree, an SVM classifier is trained using the image

processing samples from the first group as positive

instance and the sample from the second group as

negative example. Classes from the first clustering

group can be found in the first sub tree (on the left),

whereas classes from the second clustering-group can

be found in the second subtree (on the right). Once a

class per group has defined a leaf in the decision-tree,

the operation is repeated until no more classes per

group have defined leaves.

3 TESTING SETUP AND

TESTING PROCEDURE

Group = SVM classify (SVMstruct, Sample) Uses the

data from SVMStruct, a support vector machine

classifier structure created with the SVM train

function, to categorize every row of the data in Sample,

a data matrix. Similar to the training data required to

create an SVMstruct, a sample is a matrix with every

row designating an observation or duplicate and every

column designating a feature or variable. Therefore the

sample must have precisely the same number of

column as the training set. This is true because the

number of column affects the number of

characteristics. The word Group denotes the group to

which each row of sample has been assigned.

The result of A class membership is the k-NN

categorization. Before an object may be assigned to

one of its k closest classes, a majority of its

neighbours must concur on its classification (k is a

positive integer, typically small). The object is simply

given to the person who is closest when k = 1

neighbor's class. For the test set from the training set,

use k-nearest neighbor classification. For each row of

the test set, the nearest training set vectors are

identified, and the classification is determined by a

majority vote, with ties being broken at random

(Hura, Singh, and Hoe 2020). All candidates are put

up for vote the kth nearest vector, if there are ties.

3.1 Dataset

Twin information image processing technology for

these setups was gathered during data gathering

sessions during the Twins Days Festival in August

2009 in Twinsburg, Ohio. There are 186 participants

in the sample, 152 of whom are women and 34 men.

The biometric twins who participated in the data

collection assert that they are identical twins. To

formally validate the claims, no DNA testing was

done. The Human Subjects Institutional Review

Board at the University of Notre Dame authorised the

full data collection protocol for the festival (HSIRB).

3.2 Statistical Analysis

Using a statistical analysis SPSS application, the

Work is examined. In order to determine the best

method for classifying biometric real-time Indian

twins. Using SPSS, the means, standard deviations,

and mean and standard error for the SVM and LR

algorithms were calculated. Accuracy is used as the

dependent variable, and SVM and LR are treated as

the two independent variables. The sample t-test is

computed and analysed using SPSS Statistics

(Bandyopadhyay, Rout, and Satapathy 2021).

4 RESULTS

Figure 1 displays the Gabor filter's outcome. The

SVM classify method uses the input photos to

determine whether or not they are twins. The training

dataset consists of 20 images with 20 rows and

200x200 columns. After the SVM has been trained,

the training dataset is used to determine whether or

not a fresh image should be used as input. If the object

is a twin, the classification result should be 1, else it

should be -1. Fig 3 represents the SVM findings

depicts identical twins as 1 if not -1. Other biometrics

including the palm, fingerprint, iris, and speaker

identification have been used to make the distinction

between identical twins. Humans are trained to

recognise faces from birth, and there is strong

evidence that this activity takes place in the cortex's

fusiform face area. As a result, there has been a lot of

interest in creating face detection algorithms that

mimic how individuals view visual information. For

instance, face identification technology using Gabor

wavelet features with biological inspiration has

proved successful. Additionally, it has been found

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

380

that merging automatic algorithms with human

performance can greatly improve each party's

performance. The dataset is used as the input images

for identification and contains facial photographs of

identical twins taken over the course of two days in

August 2009 during the Twins Days Festival in

Twinsburg, Ohio. At random, 20 twin pairs are

chosen from the Twins Day Festival.

Table 1 clearly explains the Comparison of the

most accurate dataset's N=20 samples using the linear

SVM and LR algorithms 62.2650% and 31.0225%

respectively employing dataset size=7476, 80.5%

training data, and 19.5% test data in the sample (when

N=1).

Table 2 shows the performance of SVM and LR

on the using face and ID recognition dataset. The

results show that the mean accuracy for SVM was

62.2650, with a standard deviation of 3.56543 and a

standard error mean of 0.79725. For Logistic

Regression, the mean accuracy was 81.5815, with a

standard deviation of 4.05456 and a standard error

mean of 0.90663. In Table 3 shows the results of the

significance test, which indicates show the accuracy

of the two algorithms differs in a statistically

significant way. It is supported by the significance

value of less than p=0.408 (p<0.05) that Logistic

Regression is a superior method for this dataset and

task compared to Random Forest.

Table 3 illustrates the independent sample t-test

results produced by the significance level LR and

SVM algorithms have significant values of p = 0.408

(p<0.05). As a result, with a 95% confidence interval,

the significance threshold for both the LR and the

SVM algorithms is less than 0.05.

Table 1: Comparison between Linear SVM and LR algorithm with N=20 iteration samples of the dataset with the highest

accuracy 62.2650% and 31.0225% respectively in the sample (when N=1) using the dataset size=7172 and the 80.5% of

training & 19.5% of testing data.

Sample (N)

Dataset size/rows

SVM accuracy in %

LR accuracy in %

7182

69.78

89.33

7123

68.31

88.42

6987

67.53

87.53

6900

66.81

86.51

5087

65.23

85.23

5012

64.67

84.41

4987

63.33

83.63

4565

62.41

82.14

4444

61.30

81.90

4321

60.53

80.31

4312

60.40

79.28

4300

60.35

79.10

3099

60.26

78.98

3081

60.22

78.60

3097

59.78

78.38

3000

59.44

78.26

2098

59.38

77.68

2012

59.17

77.50

1089

58.28

77.30

1001

58.12

77.14

An Accurate Approach to Classify Real Time Indian Twins Using SVM and Compare the Performance over Logistic Regression

381

Table 2: Statistical results of SVM and LR algorithms. Mean error value, standard deviation and standard error mean for

SVM and LR algorithms are obtained for 20 iterations. It is observed that the mean for SVM (62.2650%) performed better

than the LR (31.0225%) algorithm.

Group Statistics

ACCURACY

ALGORITHMS

Mean

Std. Deviation

Std. Error Mean

SVM

62.2650

3.56543

.79725

31.0225

4.05456

.90663

Table 3: The Independent sample t-test of the significance level SVM and LR algorithms results with significant values (p <

0.05). Therefore both the SVM and the LR algorithms have a significance level less than 0.05 with a 95 % confidence interval.

Independent Samples test

Accuracy

Levene's Test for

Equality of

Variances

T-test of Equality of Means

95% of the confidence

interval of the Difference

Sig (2-tailed)

Mean

Difference

Std Error

Difference

Sig.

Lower

Upper

Equal

Variance

Assumed

.699

.408

-16.000

.001

19.31650

1.20731

-21.7605

-16.8724

Equal

Variance

Not

Assumed

-16.000

37.389

.001

19.31650

1.20731

-21.7618

-16.8711

Figure 2: Comparison of precision between the SVM algorithm and LR. The mean precision of the SVM algorithm is better

than the LR, and the standard deviation of the SVM algorithm is highly better than the LR. X-axis: SVM algorithm vs LR

Algorithm and Y-axis represents Mean Precision values 土1 SD.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

382

5 DISCUSSION

Additionally, conventional linear discriminant

analysis (LDA) and PCA-based holistic face

matching techniques are employed (LDA). The Using

the same 100 pairs as in the second experiment,

algorithms were assessed. Pittpatt and the traditional

matchmakers failed miserably at this

assignment(Khanna et al. 2021). This result is in line

with earlier research that showed how poorly face

photographs of biometric image processing identical

twins perform when subjected to face recognition

algorithms currently in use. Only the Cognitec

Matcher delivered results that were on par with

human performance and ROC. For comparison.

Additionally, It displays the average human

performance (as determined by all 25 participants).

Human observers outperform the machine matcher

practically across the whole range of False Accept

Rate, as Fi can be seen (FAR) Another drawback of

the twin study method is that if there are significant

gene-environment correlations or interactions, the

division of liability into separate genetic and

environmental components may be erroneous. Future

potential and restrictions of the image result for

indian twins You may simulate the many steps

involved in production using digital twin technology.

This is very helpful when creating a product or

manufacturing it. The virtual replica will receive real-

time data that will help fully optimise the

manufacturing process.

With a 92% accuracy rate compared to 91.32%

for logistic regression and SVM in the current system,

logistic regression is more accurate than SVM. The

suggested systems use Jupyter Notebook to train and

test the data, and SPSS Software to forecast the graph.

In the proposed system using face and ID recognition,

Logistic Regression's accuracy is anticipated to be

higher than SVM's. Using an independent dataset, the

effectiveness of logistic regression, SVM, CNN, and

other classifiers is evaluated. The assessment is

challenging because it is based on limited

data(Kyprianidis and Dahlquist 2021). When

comparing the performance of the classifiers because

it may seem easy, but it is not. The performance of

the classifiers is measured by the error rate. It is

successful if the classifier correctly categorises an

instance; otherwise, it is said to have made a mistake.

The number of data folds or partitions to utilise in this

strategy must be decided. In this study, employed a

10-fold cross-validation method, where each class is

equally represented in 10 segments of the data

(Gunjan and Zurada 2020). In this instance, the data

is split into biometric ten equal halves and used ten

times, with one tenth being used for testing and nine

tenths being used for training. Each tenth is tested

following the repeat. This enables us to estimate the

overall error using the data ten times (Tabii et al.

2018).

It has been demonstrated that the degradation of

the face is much worse than that of other biometrics,

such the iris and fingerprint. Human face recognition

ability is frequently used as a standard for evaluating

face recognition systems because people are quite

adept at identifying persons from photos (Brownlee

2019). To knowledge, no extensive human study has

been done to address the problem of telling identical

twins apart just from facial photographs. Here, test

people's ability to distinguish between facial images

of the same person and two sets of identical twin

siblings. The capacity of humans to discriminate

between identical twin brothers' faces in images

suggests that they may be able to identify features that

might be utilised to increase the face recognition

software's accuracy. In this study, volunteers rate how

certain they are that two facial photos are identical

twins or the same individual. after viewing pairs of

facial images (Huertas-Fernández et al. 2015).

6 CONCLUSION

The performance of two image processing algorithms

the current study looked into Support Vector Machine

(SVM) and Logistic Regression (LR) for

identification using face and ID recognition. The

findings clearly demonstrate that SVM has a better

accuracy of 62.2650% compared to Logistic

Regression's accuracy of 31.0225%. This shows that

SVM outperforms Logistic Regression Identification

using ID recognition as a superior and much more

suited technique.

REFERENCES

Agarwal, Basant, Valentina Emilia Balas, Lakhmi C. Jain,

Ramesh Chandra Poonia, and Manisha Sharma. (2020).

Deep Learning Techniques for Biomedical and Health

Informatics. Academic Press.

Babaeian, Mohsen, K. Amal Francis, Khalil Dajani, and

Mohammad Mozumdar. (2019). “Real-Time Driver

Drowsiness Detection Using Wavelet Transform and

Ensemble Logistic Regression.” International Journal

of Intelligent Transportation Systems Research.

https://doi.org/10.1007/s13177-019-0176-z.

An Accurate Approach to Classify Real Time Indian Twins Using SVM and Compare the Performance over Logistic Regression

383

Bandyopadhyay, Mainak, Minakhi Rout, and Suresh

Chandra Satapathy. (2021). Machine Learning

Approaches for Urban Computing. Springer Nature.

Brownlee, Jason. 2019. Deep Learning for Computer

Vision: Image Classification, Object Detection, and

Face Recognition in Python. Machine Learning

Mastery.

Deena, S. R., Kumar, G., Vickram, A. S., Singhania, R. R.,

Dong, C. D., Rohini, K., ... & Ponnusamy, V. K. (2022).

Efficiency of various biofilm carriers and microbial

interactions with substrate in moving bed-biofilm

reactor for environmental wastewater treatment.

Bioresource technology, 359, 127421.

Deshpande, A. V., Aynur Unal, Kalpdrum Passi, Dharm

Singh, Malaya Nayak, Bharat Patel, and Shafi Pathan.

(2018). Smart Trends in Information Technology and

Computer Communications: Second International

Conference, SmartCom 2017, Pune, India, August 18-

19, 2017, Revised Selected Papers. Springer.

Gaber, Mohamed Medhat, Frederic Stahl, and João Bártolo

Gomes ( 2013). Pocket Data Mining: Big Data on Small

Devices. Springer Science & Business Media.

Huertas-Fernández, I., F. J. García-Gómez, D. García-Solís,

S. Benítez-Rivero, V. A. Marín-Oyaga, S. Jesús, M. T.

Cáceres-Redondo, et al. 2015. “Machine Learning

Models for the Differential Diagnosis of Vascular

Parkinsonism and Parkinson’s Disease Using

[(123)I]FP-CIT SPECT.” European Journal of Nuclear

Medicine and Molecular Imaging 42 (1): 112–19.

Hura, Gurdeep Singh, Ashutosh Kumar Singh, and Lau

Siong Hoe. (2020). Advances in Communication and

Computational Technology: Select Proceedings of

ICACCT 2019. Springer Nature.

Khanna, Ashish, Deepak Gupta, Zdzisław Pólkowski,

Siddhartha Bhattacharyya, and Oscar Castillo. (2021).

Data Analytics and Management: Proceedings of

ICDAM. Springer Nature.

Kyprianidis, Konstantinos, and Erik Dahlquist. 2021. AI

and Learning Systems: Industrial Applications and

Future Directions. BoD – Books on Demand.

Lee, Changyong, and Daeil Kwon. (2018). “A Similarity

Based Prognostics Approach for Real Time Health

Management of Electronics Using Impedance Analysis

and SVM Regression.” Microelectronics Reliability.

Mutijah. 2018. “Logistic Regression on the Data of

Lecturer Performance Index on IAIN Purwokerto.”

Proceedings of the International Conference on

Mathematics and Islam. Ozdemir, Zeynep, and Bulent

Tugrul. 2019.

“Geofencing on the Real-Time GPS Tracking System and

Improving GPS Accuracy with Moving Average,

Kalman Filter and Logistic Regression Analysis.”

(2019) 3rd International Symposium on

Multidisciplinary Studies and Innovative Technologies

(ISMSIT).

Mahapatra, S., Vickram, A. S., Sridharan, T. B.,

Parameswari, R., & Pathy, M. R. (2016). Screening,

production, optimization and characterization of β-

glucosidase using microbes from shellfish waste. 3

Biotech, 6, 1-10.

Ptaszynski, Michal E., and Fumito Masui. (2018).

Automatic Cyberbullying Detection: Emerging

Research and Opportunities: Emerging Research and

Opportunities. IGI Global.

Ramkumar, G. et al. (2021). “An Unconventional Approach

for Analyzing the Mechanical Properties of Natural

Fiber Composite Using Convolutional Neural

Network” Advances in Materials Science and

Engineering vol. 2021, Article ID 5450935, 15 pages,

2021. https://doi.org/10.1155/2021/5450935

Reddy, V. Siva Malleswar, V. Siva Malleswar Reddy, and

T. Poovizhi. (2022). “A Novel Method for Enhancing

Accuracy in Mining Twitter Data Using Naive Bayes

over Logistic Regression.” 2022 International

Conference on Business Analytics for Technology and

Security (ICBATS).Suguna, M. Raja, M. Raja Suguna,

and P. Amaresh. 2022. “Enhanced Twitter Sentimental

Analysis Using Artificial Neural Network over Logistic

Regression towards Increase in Accuracy of

Prediction.” RECENT TRENDS IN SCIENCE AND

ENGINEERING.

Vadineanu, Serban, and Mitra Nasri. (2020). “Robust and

Accurate Period Inference Using Regression-Based

Techniques.” 2020 IEEE Real-Time Systems

Symposium (RTSS).

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

384