International Journal of Advanced Engineering Research and Science

(IJAERS)
Peer-Reviewed Journal
ISSN: 2349-6495(P) | 2456-1908(O)
Vol-11, Issue-1; Jan, 2024
Journal Home Page Available: https://ijaers.com/
Article DOI: https://dx.doi.org/10.22161/ijaers.11.2

Performance Evaluation of Neural Networks in Road Sign

Recognition
Sanjit Kumar Saha

Department of Computer Science and Engineering, Jahangirnagar University, Bangladesh

Received: 20 Nov 2023, Abstract— This paper presents an in-depth study of road sign recognition
Receive in revised form: 27 Dec 2023, techniques leveraging neural networks. Road sign recognition stands as
a critical component of intelligent transportation systems, contributing to
Accepted: 05 Jan 2024,
enhanced road safety and efficient traffic management. The paper focuses
Available online: 13 Jan 2024 on exploring various neural network architectures for example,
©2024 The Author(s). Published by AI backpropagation neural network and hybrid neural network which is a
Publication. This is an open access article under combination of two neural network (backpropagation neural network and
the CC BY license bidirectional associative memory), training methodologies, dataset
(https://creativecommons.org/licenses/by/4.0/). considerations, and performance evaluations for accurate and real-time
recognition of road signs. The experimental result shows that the hybrid
Keywords— Hybrid Neural network, Neural
neural network is faster than the backpropagation neural network in the
network, Pattern recognition, Performance
completion of the training process with higher recognition accuracy.
evaluation, Road sign recognition

I. INTRODUCTION based approaches have revolutionized this field. These

In today's dynamic and interconnected world, the safety technologies enable automated recognition and
and efficiency of transportation systems stand as paramount interpretation of road signs from images or video streams
concerns. Road sign recognition, a fundamental component captured by cameras mounted on vehicles or infrastructure.
of intelligent transportation systems, plays a pivotal role in The complexity of road sign recognition arises from
enhancing road safety, facilitating efficient traffic various factors, including diverse environmental conditions,
management, and enabling the progression towards variations in sign appearances due to aging, damage, or
autonomous driving. Recognizing and interpreting road regional differences in designs and symbols, as well as the
signs is crucial for providing timely and accurate need for real-time processing to ensure timely responses.
information to drivers and autonomous vehicles, aiding in Overcoming these challenges requires sophisticated
adherence to traffic regulations, and ensuring safe algorithms, robust training methodologies, and extensive
navigation on roadways. datasets that encompass the diversity of road signs
Road signs convey essential messages to drivers, encountered in different geographical locations and
alerting them to speed limits, warnings about hazards, environmental conditions.
providing directions, and communicating regulatory In recent years, there has been notable research
instructions. The ability to recognize these signs accurately advancement in the domain of road sign recognition.
and swiftly is imperative, as it directly influences driver Namyang and Phimoltares [1] utilized a combination of
decision-making, reduces accident risks, and contributes Support Vector Machines (SVM) and Random Forest
significantly to the overall efficiency of transportation algorithms, along with HOG and the Color Layout
networks. Descriptor (CLD), for traffic sign classification. Soni et al.
Traditionally, road sign recognition relied on manual [2] employed HOG and LBP descriptors with Principal
interpretation by human drivers. However, advancements in Component Analysis (PCA) and Support Vector Machines
computer vision, machine learning, and neural network- (SVM) for traffic sign classification. Sapijaszko et al. [3]

www.ijaers.com Page | 7
Saha International Journal of Advanced Engineering Research and Science, 11(1)-2024

proposed a traffic sign recognition system encompassing multiple neurons enables the brain to execute tasks at a
stages such as normalization, feature extraction, considerably higher speed compared to today's fastest
compression, and classification. Aziz and Youssef [4] computers.
proposed a traffic sign recognition system utilizing feature Despite their straightforward architecture, a collection
extraction and the Extreme Learning Machine (ELM) of neurons yields remarkable computational capabilities.
algorithm. Wang [5] introduced a traffic sign classification Each neuron comprises a cell body, a multitude of dendrites
system employing three machine learning classifiers: (fiber-like structures), and a solitary elongated fiber
Logistic Regression (LR), Multilayer Perceptron (MLP), identified as the axon. Dendrites form an intricate network
and SVM. surrounding the cell body, while the axon extends towards
But, neural networks, particularly Backpropagation other neurons' dendrites and cell bodies. Refer to Figure 2
Neural Network, have demonstrated remarkable capabilities for an illustrative representation of a biological neuron.
in image recognition tasks, making them a promising
approach for road sign recognition. This paper outlines the
performance evaluation of the backpropagation neural
network and the hybrid neural network.

II. PRELIMINARIES
A. Road Signs
Road signs serve as crucial elements of traffic control
and safety, providing essential information to drivers, Fig. 2 Biological neuron
pedestrians, and other road users. These signs are
standardized visual cues that communicate various
Our brain functions as an intricate and sophisticated
instructions, warnings, regulations, and guidance about road
information-processing system that operates in a highly
conditions. Figure 1 shows some road signs with their
complex, nonlinear, and parallel manner. Unlike traditional
corresponding meanings.
systems where data processing occurs in specific areas, in
neural networks, information is stored and processed
simultaneously across the entire network. This global
approach to both data and its processing distinguishes
neural networks by their widespread rather than localized
functionality. The adaptability of connections between
neurons, causing variations that contribute to arriving at the
'correct' outcome, highlights the plasticity of neural
networks. Consequently, these networks possess the
capacity to learn from experiences, marking learning as a
foundational and vital attribute of biological neural
networks. The innate ability to learn effortlessly prompted
endeavors to replicate a biological neural network's
functionality within a computer environment.
Fig. 1 Road signs C. Backpropagation Neural Network
The backpropagation neural network (BPNN)
B. Neural Networks architecture comprises an input layer, an output layer, and
one or more hidden layers. The number of input units
A neural network is characterized as a reasoning model
corresponds to the quantity of bits or values representing the
inspired by the human brain's structure. The brain is
input pattern, accommodating these 𝑛 bits or values.
composed of a highly interconnected network of nerve cells
Similarly, the count of output units is determined by the bits
known as neurons, which serve as the fundamental units for
or values associated with the output pattern.
processing information. Human brains encompass nearly
Conventionally, the network establishes complete
10 billion neurons and an extensive network of 60 trillion
connectivity exclusively between adjacent layers, forming
synapses interconnecting them, as documented by Shepherd
fully connected pathways throughout the network as shown
and Koch [6] Leveraging the simultaneous activation of
in Figure 3.

www.ijaers.com Page | 8
Saha International Journal of Advanced Engineering Research and Science, 11(1)-2024

Fig. 4 BAM Network Architecture

The main features of a BAM are given below:

Fig. 3 Backpropagation Neural Network
• A BAM comprises two layers of interconnected
neurons.
The backpropagation method uses two steps: • Neurons within one layer establish complete
1. In the ‘forward step’, the input is applied and connections with neurons in the other layer.
allowed to propagate to the output. The error • There are no interconnections among neurons
values of the output units are calculated by within the same layer.
subtracting output value from target value for each • The storage capacity and reliability of recall hinge
unit. on the network architecture and the algorithms
2. In the ‘backward step’, errors are propagated used for both recalling and learning.
backwards, and weights are modified. • Enhancing performance can be achieved by
The network's training objective involves refining the introducing additional layers or establishing more
weights to ensure that a given set of inputs yields the interconnections among neurons.
intended set of outputs. For conciseness, these input-output The input to a BAM network is a vector of real number,
combinations are often denoted as vectors. Training usually in the set {−1, +1}. The output is also a vector in
operates on the premise that each input vector aligns with a the set {−1, +1} with the same or different dimension.
target vector, symbolizing the anticipated output; These vectors can be considered as patterns, and the
collectively, these form a training pair. Typically, a network network makes heteroassociation of patterns. If the output
undergoes training with multiple such pairs. For instance, is required to be the same as input, then the network is said
an input pair might encompass a sequence of ones and zeros to make auto-association.
representing a binary image corresponding to an alphabet
letter. The compilation of these training pairs constitutes a
training set.
D. Bidirectional Associative Memory
Bart Kosko [7] introduced Bidirectional Associative
Memory (BAM) as a heteroassociative neural network as
shown in Figure 4. It operates by receiving an input pattern
represented as a vector across one group of neurons and
generates a correlated yet distinct output vector across
another set, and conversely does the same in reverse.
Fig. 5 Hybrid neural network

E. Hybrid Neural Network

A hybrid neural network consists of two distinct neural
networks: the BAM neural network and the Bidirectional
neural network. BAM is employed to reduce the dimensions
of the feature matrix, thereby enhancing the speed and

www.ijaers.com Page | 9
Saha International Journal of Advanced Engineering Research and Science, 11(1)-2024

efficiency of recognition. Figure 5 illustrates the network

architecture for this hybrid neural network.

III. EXPERIMENTS AND PERFORMANCE

EVALUATION
Recognition of Road Sign is a step-by-step processing
of road sign. These processes include:
• Preprocessing
• Recognition of sign
• Performance evaluation
Fig. 7 Image conversion
The effectiveness of the algorithm has been justified for
different Road Sign images of different resolutions. The
algorithm is capable of preprocessing and recognizing signs The road signs are arranged within a 64 × 64 grid and
of any grayscale images. The implementation of the inputted into the neural network's input layer as feature
algorithm was carried out using the C programming vectors or training patterns. Consequently, the grid
language. configurations are depicted as vectors comprising 4096
At the beginning, a sign is chosen. The sign is a components (where each vector component is 1 if the pixel
grayscale image. Whole of the task is done by the following in the grid is shaded, otherwise it is 0). The hidden layers
way: encompass 60 neurons each, which account for 60% of the
A. Preprocessing input layer. Considering a total of 64 signs, inclusive of
both mandatory and other signs, the output layer comprises
Road signs are being preprocessed by following a
64 neurons. Having 64 neurons in the output layer enables
sequence of operations: Capturing, Binary image
the representation of a 6-bit code (2^6 = 64) for classifying
conversion.
each target output. Hence, the target outputs range from
Capturing: Road sign acquisition involves utilizing a 000000 to 111111, aligning with each distinct sign.
camera for capturing purposes. For this study, specific
B. Recognition of sign
obligatory road signs were utilized to create an image
database stored as BMP type files. As part of the capturing The road sign recognition phase is most important and
process, standardization, and geometric normalization, complicated phase. Hence, the sign is 64 × 64 pixels
involving adjustments in size and direction, were applied to image. Each sign has a feature matrix of 4096 elements in
the images. For analysis purposes, the images were resized it. Each element is nothing but binary values (0 and 1).
to a resolution of 64 × 64 pixels, as depicted in Figure 6.

Fig. 6 Road sign image

Binary Image Conversion: The initial image was notably

in color. It underwent a transformation into a grayscale
image and subsequently underwent binarization, a process
chosen for its simplicity in pattern matching during sign
Fig. 8 Recognition of a sign ”NO PARKING”
recognition. Each sign is represented as a matrix of
numerical values, ranging from 0 to 255, which can be
further translated into binary format (0s and 1s). The Hence, in this experiment the number of neurons in
conversion sequence is demonstrated in Figure 7. input layer is 4096, neurons in input layer for BPNN and
output layer for BAM is 16, number of neurons in hidden
layer is 10, and finally the neurons in output layer is 6. The

www.ijaers.com Page | 10
Saha International Journal of Advanced Engineering Research and Science, 11(1)-2024

number of neurons in hidden layer can vary from 50% to 15000 1,912328 0,258816
70% of its input neurons. 20000 1,813421 0,256447
Figure 8 shows the snapshot of the program output of a 25000 1,746574 0,255095
normal image.
30000 1,698342 0,25338
C. Performance evaluation
34075 1,607801 0,009981
To assess the neural network's performance, a series of
experiments were conducted, employing separate training 40000 1,576457
and test image sets for each sign without any overlap 45000 1,53768
between them. The back-propagation neural network
50000 1,457854
underwent training utilizing default learning parameters
(learning rate 0.3, threshold 1) over 75 epochs. 55000 1,356245
Subsequently, the network was employed to recognize 60000 1,247856
individual signs.
65000 1,109854
Throughout the training process, the program continued
70000 1,065242
execution until the error reached a minimum threshold
level, illustrating the error reduction per iteration in a 75000 0,176542
graphical representation. Initially, the task was 80000 0,039947
implemented using the BPNN algorithm alone, followed by
85000 0,017469
merging the BAM and BPNN algorithms to train and
recognize road signs. Upon analysis, it was observed that 88043 0,009997
employing the hybrid network (BAM and BPNN) required
fewer iterations for training and less time for sign
recognition compared to BPNN alone. Iteration Vs. Error
Table 1 Iteration Vs. Error (70%) 5
Iteration BP (70%) BP With BAM (70%) 4
3
Error

2 3,866509 1,959864 BP
2
5 3,858203 1,894699 1
BP With
10 3,503066 1,881338 0
BAM
0 50000 100000
100 3,435648 1,702669
Iteration
500 3,217658 1,119279
1000 3,095643 0,612061 Fig. 9 Iterations Vs. Error (70%)
1500 2,943532 0,575292
2000 2,873423 0,554884 The training process involved eight training input
2500 2,806753 0,543144 patterns and employed an error threshold (e.g., 0.001) to
halt training. The number of iterations was contingent upon
3000 2,645987 0,53713
the percentage of the hidden layer and the algorithm utilized
3500 2,546534 0,531881 during training. For instance, setting the hidden layer to
4000 2,513423 0,492877 70% of the input layer resulted in 88043 iterations, while
reducing the hidden layer to 50% led to 43067 iterations.
4500 2,485645 0,315097
Upon adopting the hybrid network with the same
5000 2,437864 0,293473
percentage of the hidden layer, the iteration count decreased
6000 2,376588 0,278986 further. Specifically, it decreased from 34075 to 14977,
7000 2,238757 0,272395 illustrating an even more pronounced reduction in iterations
compared to previous results.
8000 2,187658 0,268759
9000 2,074542 0,266053
10000 1,984532 0,264022

www.ijaers.com Page | 11
Saha International Journal of Advanced Engineering Research and Science, 11(1)-2024

Table 2 Iteration Vs. Error (50%) Comparing the performance of two networks as BPNN
Iteration BP (50%) BP With BAM (50%) and hybrid neural network a decision had been taken that
the hybrid neural network takes less iteration than BPNN in
2 6,933186 3,07014 completion of the training process.
5 5,694922 3,055324
10 4,678283 3,025566 IV. CONCLUSION
100 4,333929 1,374816 Neural network-based road sign recognition holds
500 4,045357 0,567401 immense potential in revolutionizing transportation systems
by enhancing road safety and traffic management. The
1000 3,874532 0,316263
paper underscores the significance of neural networks in
1500 3,523548 0,134039 this domain and outlines the performances of BPNN and
2000 3,223324 0,077653 hybrid neural networks with experimental results. And the
result shows that the hybrid neural network performs faster
2500 2,963092 0,057539
than BPNN with high recognition accuracy.
3000 2,927669 0,04723
3500 2,906451 0,040717
REFERENCES
4000 2,566751 0,035315 [1] Namyang, N. & Phimoltares, S. (2020). Thai traffic sign
4500 2,550943 0,031443 classifica- tion and recognition system based on histogram of
gradients, color layout descriptor, and normalized correlation
5000 2,539485 0,028615 coefficient. International Conference on Information
6000 2,530225 0,023834 Technology (InCIT), Chonburi, Thailand, 270-275.
[2] Soni, D., Chaurasiya, R.K. & Agrawal, S. (2019). Improving
7000 2,522263 0,020596 the Classification Accuracy of Accurate Traffic Sign
8000 2,515119 0,018211 Detection and Recognition System Using HOG and LBP
Features and PCA-Based Dimension Reduction. Proceedings
9000 2,502996 0,016202
of the International Conference on Sustainable Computing in
10000 2,448071 0,014673 Science, Technology and Management (SUSCOM), Amity
University Rajasthan, Jaipur, India.
14977 2,056355 0,01
[3] Sapijaszko, G., Alobaidi, T. & Mikhael, W.B. (2019). Traffic
20000 1,800224 sign recognition based on multilayer perceptron using DWT
and DCT. Proceedings of the 2019 IEEE 62nd International
25000 1,034644
Midwest Symposium on Circuits and Systems (MWSCAS),
30000 0,176542 Dallas, TX, USA, 440-443.
[4] Aziz, S. & Youssef, F. (2018). Traffic sign recognition based
35000 0,039947
on multi-feature fusion and ELM classifier. Procedia
40000 0,017469 Computer Science, 127, 146-153.
[5] Wang, B. (2022). Research on the Optimal Machine Learning
43067 0,009997
Classifier for Traffic Signs. Web of Conferences; EDP
Sciences: Les Ulis, France, 144, 03014.
[6] Shepherd, G. M. & Koch, C. (1990). Dendritic electrotonus
Iteration Vs. Error and synaptic integration. The Synaptic Organization of the
Brained, G. M. Shepherd, Oxford University Press, appendix.
10 [7] Kosko, B. (1988). Bidirectional associative memories. IEEE
Error

BP Transactions on Systems, Man, and Cybernetics, 18(1), 49-

5
60.
0 [8] Saha, S.K., Chakraborty, D. & Bhuiyan, Md. Al.Amin
BP With (2012). Neural Network based Road Sign Recognition.
0 20000 40000 60000 BAM International Journal of Computer Applications, 50, 35-41.
Iteration

Fig. 10 Iterations Vs. Error (50%)

www.ijaers.com Page | 12