Improving the Accuracy of Identifying Real-Time Indian Twins

Using CNN Compared with Random Forest

Vallipi Dasaratha

and J. Joselin Jeya Sheela

†

Department of Electronics and Communication Engineering, Saveetha School of Engineering, Saveetha Institute of Medical

and Technical Sciences, Saveetha University, Chennai, Tamilnadu, 602105, India

Keywords: Biometric, Convolutional Neural Network, Face Recognition, Identification, Image Processing, Novel,

Technology.

Abstract: The objective of this study is to achieve real-time identification and analysis of Indian twins using the random

forest algorithm, while also comparing its performance with the Convolutional Neural Network (CNN)

algorithm in terms of accuracy. Materials and Methods: For the purpose of face recognition of twins with face

and ID recognition, the random forest algorithm is chosen over the Convolutional Neural Network (CNN).

The study involves two groups, namely Group 1 and Group 2, with an overall sample size of 1430 and 20

sample iterations for each group. Results and Discussion: The comparison and classification of real-time

Indian twins are conducted using the Random Forest algorithm and the performance is measured using the

CNN algorithm. The achieved accuracy rates are 52.3965% for Random Forest and 64.305% for CNN. By

comparing the accuracy of both algorithms, independent sample tests reveal a statistically significant

difference with a p-value of 0.001 (p<0.05), confirming the significance of the hypothesis through an

independent sample t-test. Conclusion: This study evaluated the effectiveness of two image processing

algorithms, namely Random Forest and CNN. The results indicate that Random Forest achieves an accuracy

of 52.3965%, outperforming CNN which achieved an accuracy of 64.3050%. This suggests that for

identification using ID recognition, Random Forest provides superior performance compared to CNN.

1 INTRODUCTION

There are diverse applications for this intriguing

challenge, including social media photo tagging,

activity tracking, crime detection, and more. It poses

a intricate visual challenge (FGVC) due to the minute

inter-class variances of the twin objects. Owing to

their remarkable accuracy, they are frequently

employed (Chandana.S, Harini, and Senthil 2022).

The utilization of object detectors like Yolo3, Faster-

RCNN, SSD, and ResNet-101, alongside pretrained

base networks such as VGGNet, ResNet-101,

Inception with ResNet, and Retina Net, has shown

promising outcomes. Nevertheless, the accuracy of

all these models diminishes when objects with

exceedingly slight differences need to be recognized.

CNNs acquire significant low-level features in a

hierarchical, feed-forward manner, which may lead to

smoother learning progression (Chandana.S, Harini,

and Senthil 2022; Lane et al. 2015), impacting the

Research Scholar

†

Research Guide, Corresponding Author

model's adeptness in detecting fine-grained objects.

The incorporation of features taught at diverse levels

and scales is essential for this objective. To avert

overfitting, a substantial dataset is required. We

devised a solution by employing publicly available

images of renowned twins Mary-Kate and Ashley

Olsen, generating an annotated dataset of 120 images

to address my twin identification challenge (Shoji and

Zhang 2019). This dataset comprises images of the

twins standing alongside other individuals, singly or

together, constituting my dataset's principal

challenge. To surmount this, I employed a VGGNet

CNN trained with the ImageNet dataset, and applied

the Single Shot Detector (SSD) based on the trained

VGGNet base network. SSD is a rapid one-pass

detector (B402. et al. 2021) with low computational

demands suitable for video detection. Despite a

marginal trade-off in accuracy when objects are very

small, SSD excels in recognizing them by capturing

features across diverse scales. This study's primary

Dasaratha, V. and Sheela, J.

Improving the Accuracy of Identifying Real-Time Indian Twins Using CNN Compared with Random Forest.

DOI: 10.5220/0012520100003739

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 487-493

ISBN: 978-989-758-661-3

487

contributions encompass photographing the "Olsen"

twins in their natural environment to comprehend

twin identification complexities and providing an

effective method for fine-grained item detection in

scenarios with limited datasets. Detailed exploration

of current fine-grained object detection research is

covered in the second part of this article (Gharbi et al.

2018; Hegde and Manjunatha 2022). The third and

fourth sections describe the dataset and model,

respectively, while Section 5 presents the research

findings.

In my scenario, there exist exceedingly subtle

distinctions between the two distinct groups. My twin

subjects, referred to as "Mary" and "Ashley,"

represent the classes. These classes exhibit extremely

delicate dissimilarities that render their

differentiation challenging across all images

(Mahapatra, S et al. 2016). The novel learning model

must diligently capture these distinguishing attributes

to the maximum extent possible. Moreover, it should

not hastily incorporate irrelevant features, as this

would obscure the significance of discriminative

characteristics in the detection process. Employing

transfer learning in CNN models expedites training

while augmenting novel learning, particularly when

the dataset is constrained (Vinod and Padmapriya

2022). Through pre-training a model on a relevant

dataset, it can be adapted to a different task using a

smaller dataset for that purpose. Despite employing

distinct datasets for the two tasks, the model's weights

can be adjusted. Employing a slightly different

approach, I extended this notion to address the twins'

detection challenge. My dataset encompasses images

with and without immediately distinguishable core

attributes. Following the initial update of my pre-

trained model with the non-essential images, only the

final layer was modified using the critical subset. This

approach intuitively aligns with CNN learning,

allowing the utilization of the most distinctive

attributes within the upper layers (DeVerse and Maus

2016).

2 MATERIALS AND METHODS

The experimentation took place within the Machine

Learning Laboratory at Saveetha School of

Engineering, Saveetha Institute of Medical and

Technological Sciences. Two distinct entities were

established for the study: Group 1 denoting CNN and

Group 2 corresponding to Random Forest.

Employing a G-power of 80%, the system calculates

the required sample size and defines it as 40 iteration

samples while accessing the Clincale website (Group

1 - 20, Group 2 - 20). The setup consists of two

separate groups, with a cumulative sample size of

1430. Each of the two groups, Group 1 and Group 2,

underwent 20 sample iterations.

The sample preparation process for the renowned

Random Forest machine learning technique within

the context of the supervised technology learning

novel approach has been completed. This approach

holds potential advantages for machine learning tasks

encompassing both classification and regression. It is

rooted in the innovative ensemble learning theory, a

strategy that integrates multiple classifiers to address

challenging problems and enhance the overall model

performance.

Sample preparation for Group 2, focused on CNN,

has been completed. CNN, as a classification

algorithm, accomplishes the task by transforming the

initial training data into a higher-dimensional space

through a nonlinear mapping. Within this transformed

space, Random Forest seeks an optimal hyperplane

that effectively separates instances of distinct classes.

The determination of this hyperplane is influenced by

the cases that closely interact with the division

between the two classes, known as support vectors.

The effectiveness of the Random Forest classifier is

significantly influenced by the chosen kernel function

and its associated parameters.

The most favorable outcomes are often achieved

by employing a Puk kernel along with a kernel

parameter value of C = 1.

The development of the face recognition

identification system was carried out using Jupyter

Notebook on a Windows 11 operating system. The

system's implementation involves two distinct

groups: Random Forest and CNN methods. These

algorithms are integrated into a novel dataset, which

is then subjected to training and testing processes to

enhance accuracy. The sample dataset consists of 40

instances. During the model training, various loss

functions were employed. To better align with the

correct labels, the initial cross-entropy loss was

adjusted to focal loss, as inspired by the SSD study

investigation.

The cross-entropy loss balances the weights of

both positive and negative instances, but it doesn't

differentiate between simple and complex cases. In

contrast, focal loss reshapes the cross-entropy loss by

reducing the weight applied to well-classified or

simple data. The focal loss function is explained for

classification, with "alpha" representing the

balancing parameter and "gamma" representing the

focusing parameter.

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

and Consumer Electronics

488

3 RANDOM FOREST

The widely recognized Random Forest is a machine

learning algorithm that employs an innovative

ensemble learning technique for prediction purposes.

It can be employed in supervised learning scenarios

for both classification and regression tasks. The

algorithm's functionality involves training numerous

individual classifiers, referred to as decision trees,

and subsequently amalgamating their predictions to

formulate a final prediction. This approach proves

efficacious in enhancing the model's performance and

addressing intricate challenges.

Algorithm for random forest

Step1:from sklearn.datasets import

fetch_olivetti_faces

Step2: from sklearn.model_selection import

train_test_split

Step3: from sklearn.metrics import

accuracy_score

Step4: from sklearn.ensemble import

RandomForestClassifier

Step5:X-

data.reshape((data.shape[0],data.shape[1]*data.shap

e[2]))

Step6:X_train, X_test, y_train,

y_test=train_testsplit(X,target, test_size=0.3,

stratify=target, random_state=0)

Step7:clf = RandomForestClassifier()

Step8:clf.fit(X train pca, y train)

Step9:y pred-clf.predict(X_test_pca)

Step10: print(accuracy_score(y_pred,y_test)*

100)

3.1 Convolutional Neural Network

(CNN)

Convolutional Neural Networks (CNNs), often

known as convnets or CNNs, constitute a crucial facet

of machine learning. They represent a subset of

artificial neural network architectures that find

application in diverse and pioneering objectives and

datasets. Specifically tailored for deep learning

algorithms, CNNs serve as a network architecture

designed for image recognition and the processing of

pixel data.

4 STATISTICAL ANALYSIS

The investigation was conducted using IBM SPSS

version 21. The research independent factors

encompassed the project, V name, and year end.

Meanwhile, the research dependent variables

involved face and face ID. Iterations with a limit of

15 instances were executed for both the proposed and

existing methods. The anticipated accuracy for each

innovative iteration was recorded for accuracy

analysis. Subsequently, the results of these iterations

were subjected to an Independent Sample T-test. A p-

value of p=0.350 (p<0.05) indicated that there was no

discernible disparity in the accuracy of the algorithms

("Real-Time Modeling of Albedo Pressure on

Spacecraft and Applications for Improving

Trajectory Est." 2023).

5 DATASET

A dataset of widely available photos featuring the

Olsen twins was meticulously assembled. Employing

a PythonScript, images were scraped from Google

Images, subsequently annotated using the Labeling

tool. In total, 120 photos were amassed, out of which

50 were allocated for testing purposes, and 800 were

designated for training and validation. The classes

were denoted as "mary" and "ashley." Annotating the

images presented a challenge due to the fact that

numerous photographs lacked clear facial distinctions

between the twins.

6 RESULTS

Table 1 clearly indicates that Random Forest

outperformed CNN significantly in the context of

identification using face and ID recognition. The

precision and performance of Random Forest

surpassed those of CNN, signifying its superiority for

this specific dataset and task-

Table 2 presents the efficacy of CNN and Random

Forest on a face and ID recognition dataset. The

outcomes demonstrate that Random Forest achieved

a mean accuracy of 52.3965, accompanied by a

standard deviation of 1.49933 and a standard error

mean of 0.33526. In contrast, CNN exhibited a mean

accuracy of 64.3050, with a standard deviation of

1.34707 and a standard error mean of 0.30121.

Improving the Accuracy of Identifying Real-Time Indian Twins Using CNN Compared with Random Forest

489

Table 1: Comparison between Random Forest and CNN with N=20 iteration samples of the dataset with the highest accuracy

72% and 66% respectively in the sample (when N=1) using the dataset size=7476 and the 66.45% of training & 62.30% of

testing data.

Sample (N)

Dataset Size / rows

Random Forest

accuracy in %

CNN

accuracy in %

7182

72.15

66.66

7123

72.10

66.45

6987

72.05

66.10

6900

71.98

65.89

5087

71.87

65.50

5012

71.77

65.28

4987

71.56

65.18

4565

71.45

64.50

4444

71.18

64.38

4321

70.67

64.28

4312

70.48

63.96

4300

70.34

63.86

3099

70.28

63.67

3081

70.16

63.54

3097

69.76

63.23

3000

69.46

63.78

2098

68.89

63.78

2012

68.47

62.72

1089

67.66

62.60

1001

66.45

62.30

Table 2: Statistical result of Random Forest algorithm and CNN algorithm. Mean error value, SD and standard error mean

for RF and CNN algorithm are obtained for 20 iterations. It is observed that the mean for Random Forest (52.3965%)

performed better than the CNN (64.305%) algorithm.

Group Statistic

ACCURACY

ALGORITHMS

Mean

Std. Deviation

Std. Error Mean

Random Forest

52.3965

1.49933

.33526

CNN

64.3050

1.34707

.30121

In Table 3, the results of the significance test

reveal a substantial distinction in the accuracy of the

two algorithms. The significance value of less than

p=0.350 ( p<0.05) underscores the preference for

CNN over Random Forest for this dataset and task.

Figure 1 graphically portrays the mean accuracy

of identification using face and ID recognition for both

Random Forest and CNN. The depicted results

underscore that Random Forest attained an accuracy

of 52.3965%, whereas CNN achieved an accuracy of

64.3050%. This underlines CNN's better performance

compared to Random Forest on this dataset and task.

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

and Consumer Electronics

490

Table 3: The independent sample t-test of the significance level Random Forest and CNN algorithms results with significant

values (p < 0.05). Therefore, both the Random Forest and the CNN algorithms have a significance level less than 0.02 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

Differ-ence

Std Error

Differen-ce

Sig.

Equal

Variance

Assumed

.103

.350

13.699

0.01

6.17400

.45070

5.28181

7.08639

Equal

Variance

Not

Assumed

13.699

37.572

0.01

6.17400

.45070

5.28127

7.08673

Figure 1: Comparison of precision between the CNN algorithm and RF The mean precision of the CNN algorithm is better

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

Algorithm and Y-axis represents Mean Precision values 土2 SD.

7 DISCUSSION

The aforementioned research study demonstrated that

Random Forest achieved a higher accuracy of

52.3965% as compared to CNN's accuracy of

64.3050%. A statistically significant difference

between the accuracy of the two algorithms was

determined through the utilization of independent

sample t-tests, yielding a p-value of p=0.350 (p<0.05)

(Nasri and Kargahi 2012). The dataset employed for

this study was sourced from the open-source platform

Kaggle and was applied for identification through

face and ID detection. In the current system, Random

Forest displayed superior accuracy to CNN,

achieving 52.3965% accuracy in contrast to CNN's

64.3050%. For the proposed system, the dataset was

trained and tested using applications such as SPSS

and Jupyter Notebook, which were also used for

forecasting graphs (DeVerse and Maus 2016).

In the proposed system for fake voter

identification using face and ID recognition, it is

anticipated that Random Forest will exhibit higher

accuracy compared to CNN. The performance of

various classifiers including Random Forest, KNN,

Improving the Accuracy of Identifying Real-Time Indian Twins Using CNN Compared with Random Forest

491

CNN, etc., is evaluated using an independent dataset,

a task that presents challenges due to limited available

data (Yadav and Kumar, n.d.; Gao, Xu, and Wang

2003). Assessing classifier performance is intricate

when comparing different learning methods, as it

involves evaluating the error rate, which determines

the classifier's success in correctly categorizing

instances (Agarwal et al. 2020). This evaluation is

achieved by considering the mistakes made by the

classifier in each instance. To effectively gauge

classifier performance, independent test data not used

in the model is employed. If additional data is

required, it can be partitioned into training and testing

sets.

Increasing the volume of training data leads to

higher classification accuracy and enhances the utility

of testing data. Nevertheless, a challenge emerges

when the available data is insufficient. To address

this, manual separation of training and test data is

essential (Samek et al. 2019) (Ramkumar, G. et al.

2021). Insufficient data can also introduce issues. To

mitigate this, the holdout approach is commonly

employed, allocating one-third of the data for testing

and the remaining two-thirds for analysis. Cross-

validation is another effective strategy, necessitating

a decision on the number of data folds or partitions to

utilize. In this research, a 10-fold cross-validation

method was adopted, splitting the data into ten

segments with equal representation across classes

(Gunjan and Zurada 2020). This approach involves

dividing the data into ten equal parts and iteratively

using 10% for testing and 90% for training. After

each iteration, one tenth is designated for testing. This

process allows for estimating the overall error over

ten iterations (Khanna et al. 2021).

A notable limitation of the twin study research

method lies in the potential influence of significant

gene-environment correlations or interactions. Such

factors can introduce inaccuracies when attempting to

segregate liability into distinct genetic and

environmental components. In the realm of

technology, a parallel concept to twins emerges

through the utilization of Indian twin technology.

This concept, prevalent within the industrial sector,

involves creating digital replicas of objects or

processes. To achieve this, sensors are strategically

positioned to collect real-time data from the physical

process, which is then fed into AI systems for

processing. Subsequently, these digital twins offer a

platform to comprehensively examine and simulate

the operational mechanics of the object or process,

facilitating in-depth insights into product behavior

and performance simulations.

8 CONCLUSION

The research study focused on the evaluation of two

image processing algorithms, Random Forest and

CNN, for the purpose of identification using face and

ID recognition. The results revealed that Random

Forest exhibited a higher accuracy of 52.3965% in

contrast to CNN's accuracy of 64.3050%. These

findings signify that Random Forest outperforms

CNN in the realm of ID recognition-based

identification.

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.

B402., Escuela Politécnica Superior Uam, Escuela

Politécnica Superior Uam, B402., and Hadi Abooei

Mehrizi. (2021). “Identifies Polyps in Real Time with

Accuracy 96.67% in Screening Colonoscopy Using

Convolutional Neural Networks (CNN).” IBJ Plus.

Chandana.S, Harini, Chandana S. Harini, and Kumar R.

Senthil. (2022). “A Deep Learning Model to Identify

Twins and Look Alike Identification Using

Convolutional Neural Network (CNN) and to Compare

the Accuracy with SVM Approach.” ECS Transactions.

https://doi.org/10.1149/10701.14109ecst.

DeVerse, Shawn, and Stefan Maus. (2016). “Improving

Directional Survey Accuracy through Real-Time

Operating Centers.” Day 2 Tue, August 23, 2016.

https://doi.org/10.2118/180652-ms.

Gao, Jingbo, Minqiang Xu, and Rixin Wang. (2003).

“Study About Real-Time Finite Element Method Using

CNN.” Computer Technology and Applications.

https://doi.org/10.1115/pvp2003-1908.

Gharbi, Salem Al, Salem Al Gharbi, Qinzhuo Liao,

Salaheldin Elkatatny, and Abdulazeez Abdulraheem.

(2018). “IIncreasing ANN Accuracy, by Improving the

Training Dataset Criteria. Case Study: Identify the

Formation Density from The Drilling Surface

Parameters in Real-Time.” All Days.

https://doi.org/10.2118/192363-ms.

Gunjan, Vinit Kumar, and Jacek M. Zurada. (2020).

Proceedings of International Conference on Recent

Trends in Machine Learning, IoT, Smart Cities and

Applications: ICMISC 2020. Springer Nature.

Hegde, Chandan R., and H. T. Manjunatha. (2022). “Real

Time Indian Traffic Sign Detection Using Image

Processing and CNN.” International Journal of

Advanced Research in Science, Communication and

Technology. https://doi.org/10.48175/ijarsct-5363.

Khanna, Ashish, Deepak Gupta, Zdzisław Pólkowski,

Siddhartha Bhattacharyya, and Oscar Castillo. (2021).

Data Analytics and Management: Proceedings of

ICDAM. Springer Nature.

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

and Consumer Electronics

492

Lane, D., S. Hill, J. Huntingford, P. Lafuente, R. Wall, and

K. Jones. (2015). “Effectiveness of Slow Motion Video

Compared to Real Time Video in Improving the

Accuracy and Consistency of Subjective Gait Analysis

in Dogs.” Open Veterinary Journal.

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.

Nasri, Mitra, and Mehdi Kargahi. (2012). “A Method for

Improving Delay-Sensitive Accuracy in Real-Time

Embedded Systems.” 2012 IEEE International

Conference on Embedded and Real-Time Computing

Systems and Applications.

https://doi.org/10.1109/rtcsa.2012.39.

Ramalakshmi, M., & Vidhyalakshmi, S. (2021). GRS

bridge abutments under cyclic lateral push. Materials

Today: Proceedings, 43, 1089-1092.

Paramasivam, G., Palem, V. V., Sundaram, T., Sundaram,

V., Kishore, S. C., & Bellucci, S. (2021).

Nanomaterials: Synthesis and applications in

theranostics. Nanomaterials, 11(12), 3228.

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

(“Real-Time Modeling of Albedo Pressure on Spacecraft

and Applications for Improving Trajectory Est.”

(2023). https://doi.org/10.2514/6.2023-2207.vid.

Samek, Wojciech, Grégoire Montavon, Andrea Vedaldi,

Lars Kai Hansen, and Klaus-Robert Müller. (2019).

Explainable AI: Interpreting, Explaining and

Visualizing Deep Learning. Springer Nature.

Shoji, Yuta, and Lifeng Zhang. (2019). “Research for

Improving Identification Accuracy of Specific Fish

Species with CNN.” Proceedings of The 7th

International Conference on Intelligent Systems and

Image Processing 2019.

https://doi.org/10.12792/icisip2019.036.

Vinod, G., and G. Padmapriya.(2022). “An Adaptable Real-

Time Object Detection for Traffic Surveillance Using

R-CNN over CNN with Improved Accuracy.” 2022

International Conference on Business Analytics for

Technology and Security (ICBATS).

https://doi.org/10.1109/icbats54253.2022.9759030.

Yadav, Anu, and Ela Kumar. n.d. “Instance Segmentation

for Real-Time Video Detection Using FPN and Mask

R-CNN.” Zohourian, Farnoush, Borislav Antic, Jan

Siegemund, Mirko Meuter, and Jose

Pauli. (2018). “Superpixel-Based Road Segmentation for

Real-Time Systems Using CNN.” Proceedings of the

13th International Joint Conference on Computer

Vision, Imaging and Computer Graphics Theory and

Applications.

Improving the Accuracy of Identifying Real-Time Indian Twins Using CNN Compared with Random Forest

493