Improving the License Plate Character Segmentation Using Na
¨
ıve
Bayesian Network
Abdenebi Rouigueb
1 a
, Fethi Demim
2 b
, Hadjira Belaidi
4 c
, Ali Zakaria Messaoui
3 d
,
Mohamed Akrem Benatia
1 e
and Badis Djamaa
1
1
Artificial Intelligence and Virtual Reality Laboratory, Ecole Militaire Polytechnique, Bordj El Bahri, Algiers, Algeria
2
Guidance and Navigation Laboratory, Ecole Militaire Polytechnique, Bordj El Bahri, Algiers, Algeria
3
Complex Systems Control and Simulation Laboratory, Ecole Militaire Polytechnique, Bordj El Bahri, Algiers, Algeria
4
Signals and Systems Laboratory, Institute of Electrical and Electronic Engineering,
University M’hamed Bougara of Boumerdes, Boumerdes, Algeria
Keywords:
License Plate, Character Segmentation, Na
¨
ıve Bayesian Network, DTW, CNN.
Abstract:
Character segmentation plays a pivotal role in automatic license plate recognition (ALPR) systems. Assuming
that plate localization has been accurately performed in a preceding stage, this paper mainly introduces a
character segmentation algorithm based on combining standard segmentation techniques with prior knowledge
about the plate’s structure. We propose employing a set of relevant features on-demand to classify detected
blocks into either character or noise and to refine the segmentation when necessary. We suggest using the na
¨
ıve
Bayesian network (NBN) classifier for efficient combination of selected features. Incrementally, one after
one, high computational cost features are computed and involved only if the low-cost ones cannot decisively
determine the class of a block. Experimental results on a sample of Algerian car license plates demonstrate
the efficiency of the proposed algorithm. It is designed to be more generic and easily extendable to integrate
other features into the process.
1 INTRODUCTION
ALPR systems (Du et al., 2013) have become crucial
in various real-world applications, such as parking
access control, road traffic monitoring, and tracking
stolen vehicles. A typical ALPR system comprises
four main modules: image acquisition, license plate
localization (PL), character segmentation (CS), and
character recognition (CR)(Du et al., 2013).
Compared to PL and CR, most recognition er-
rors at the ALPR pipeline end occur in CS due to
difficulties in correctly partitioning the plate image
into blocks containing individual characters (He et al.,
2008; Pan et al., 2008). The segmentation module
takes the filtered-normalized plate image as input and
then attempts to determine the smallest image seg-
a
https://orcid.org/0000-0001-5699-2721
b
https://orcid.org/0000-0003-0687-0800
c
https://orcid.org/0000-0003-2424-626X
d
https://orcid.org/0000-0001-5753-5776
e
https://orcid.org/0000-0003-1779-2705
ments expected to contain a single character each.
For a comprehensive recent review of CS methods
and their types in ALPR systems, refer to (Mufti and
Shah, 2021; Du et al., 2013).
The three most popular existing CS approaches
are horizontal-vertical projection (Jagannathan et al.,
2013), connected component analysis (CCA) (Yoon
et al., 2011), and character detection using deep
neural networks which represent the state-of-the-art
trend. For the last approach, the convolutional neu-
ral networks (CNN) architectures are the most used
for CS. For instance, the models YOLO (Redmon
et al., 2016), Faster-RCNN (Ren et al., 2015), and
Classification-Regression Network (CR-NET) (Mon-
tazzolli and Jung, 2017) are used for character detec-
tion and recognition in real-world ALPRs systems. It
is well to point out that all these methods are flawed
and not universal. Furthermore, each method has its
limitations, advantages, and appropriate conditions of
use. See (Mufti and Shah, 2021) for exploring their
respective performance summary, pros, and cons.
For some real conditions, these standard CS tech-
Rouigueb, A., Demim, F., Belaidi, H., Messaoui, A., Benatia, M. and Djamaa, B.
Improving the License Plate Character Segmentation Using Naïve Bayesian Network.
DOI: 10.5220/0012091500003543
In Proceedings of the 20th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2023) - Volume 2, pages 61-68
ISBN: 978-989-758-670-5; ISSN: 2184-2809
Copyright © 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
61

niques are not accurate enough when applied solely
because they do not profit from the entire prior in-
formation about the license plate and the characters’
structures. Indeed, for instance, characters in ALPR
systems have almost the same size; lie on one or two
straight horizontal layers; and pertain to a finite alpha-
bet set, etc. Therefore, sophisticated ALPR systems
in practice are those that can improve significantly CS
by involving, if necessary for difficult cases, a part of
the prior information. To achieve this goal, we need
to precisely evaluate the confidence (i.e. probability)
that a detected block corresponds to a useful charac-
ter, then add the appropriate post-processing only for
the confused blocks. Doing so helps to optimize the
trade-off between the accuracy and the time of CS. In
this context, this paper addresses the character seg-
mentation problem by combining one of these stan-
dard techniques with license plate preliminary infor-
mation using an NBN classifier.
The paper is structured as follows. The next sec-
tion introduces the NBN classifier. Then, Section 3
describes the proposed algorithm and presents some
suggested features, such as the dynamic time-warping
(DTW) distance. In this section too, we show and
discuss the obtained results. Finally, we comment on
some conclusions, and we state future works.
2 NA
¨
IVE BAYESIAN NETWORK
CLASSIFIER
NBN classifier (Mittal et al., 2007) is a simple and
powerful Bayesian network. Even with the restric-
tive independence assumption among the features, it
yields competitive performance against other state-of-
the-art classifiers. When NBN is used for classifica-
tion, feature variables x
i
are assumed to be condition-
ally independent given the class variable C. Its re-
spective structure is shown in Fig. 1, in which the
conditional joint probability P(x
i
,x
j
|C) can be factor-
ized into a product as follows:
P(x
i
,x
j
|C) = P(x
i
|C)P(x
j
|C), 1 ≤ i ̸= j ≤ n. (1)
NBN classifier predicts a new data point as the
class with the highest posterior probability using the
following formula:
ω = argmax
c
i
P(C = c
i
)
n
∏
j=1
P(x
j
|C = c
i
). (2)
Bayesian networks, including NBNs, can deal im-
mediately with missing features (not yet calculated).
By using only the available features set Y ⊂ X =
{x
1
,...,x
n
} (X all considered features), the classifica-
tion rule is as follows:
Figure 1: Na
¨
ıve Bayesian network classifier structure.
ω = argmax
C
i
P(C = c
i
)
∏
x
j
∈Y
P(x
j
|C = c
i
). (3)
To construct the NBN, we need to discretize all
features’ variables. The distributions P(C) and
P(x
i
|C) i = 1..n can be fixed by the user or estimated
using the training data.
3 PROPOSED SOLUTION
3.1 Segmentation Algorithm
Our main contribution consists of performing a stan-
dard CS technique such as CCA, projection, or CNN.
Then, NBN is incrementally utilized on request to re-
fine the CS split positions using prior information by
computing some specific features. The global scheme
of the proposed algorithm is depicted in Fig. 2. Since
our work focuses only on improving CS, we per-
formed PL by manually double-clicking on the four
edges of each plate. Hence, all eventual errors will
be due to the CS stage, not plate extraction. In the
pre-processing phase (step (a) of the algorithm shown
in Fig. 2), a geometrical normalization is carried out
by finding an affine transformation (Marcel, 1987)
from the localized plate I
1
to a 30 × 140 pixels hor-
izontal fixed-size non-tilted plate I
2
. Indeed, we have
I
2
= A × I
1
+ B where the transformation parameters,
the 2 × 2 matrix A and the 1 ×2 vector column B, can
be estimated using only three among the four edge
correspondences. Since the distance separating the
camera from the car plate is so important compared
to the car plate size, straight lines seem to remain
straight, and parallel lines seem to remain parallel,
but rectangles may change into parallelograms. It’s
worth remembering that affine transformation is suit-
able for such a kind of image shearing. After that,
pixel intensity is normalized into the range [0 − 255]
grey color to reduce the effect of the variation in light-
ing conditions. Then, the Sauvola adaptive threshold
(Sauvola and PietikaKinen, 2000) is used for bina-
rization. Next, a morphological opening is applied
to remove small objects from the image while pre-
serving the shape and the size of large ones. Finally,
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
62

simple horizontal and vertical projections can be used
to eliminate the eventual remaining borders resulting
from the lack of precision in the PL step. It is worth
noting that binarization is not recommended (may be
skipped) if deep neural networks are used for charac-
ter detection in the next step.
In step (b) of the algorithm shown in Fig. 2, we
propose to perform an initial segmentation using one
of the state-of-the-art CS methods such as CCA, pro-
jection, or deep neural networks.
Objects (blocks) found at this level can correspond
to either characters or non-characters (noise decora-
tive patterns, etc). Prior information, such as the
shape, size, position, etc., of characters in license car
plates is necessary to confirm reliably the class (Char-
acter or non-character) of each block.
In this study, for illustrative purposes only, we
have proposed to use the 07 following features of each
detected block B: x
1
= Height(B), x
2
= width(B),
x
3
= MeanPixelIntensity(B), x
4
= allign(B) (the
number of blocks horizontally aligned with B), x
5
=
HeightCon f irmedCh(B), x
6
= min(2 − DDW (B,λ
i
))
with λ
i
∈ {templates of the characters’:’0’,..., ’9’},
and x
7
= CNN(B) (the classification score of B into
character|non-Character classes). x
5
is equal to the
normalized difference between the height of B and the
mean height of the blocks already confirmed as char-
acters. x
7
is computed using a CNN model (the model
VGG16 is used in our case).
Previously, before running this algorithm, the do-
main of features is discretized such that values within
the same discrete value contribute roughly in the same
manner in prediction. For example, for x
1
, the three
proposed values 1,2 and 3 correspond to the intervals
[18 25], [14 17] + [26 29], and [0 13] + [30 + ∞] ,
respectively. Most characters have x
1
= 1, and most
non-characters have x
1
= 3. Then conditional proba-
bility tables P(x
i
/C) are estimated using training la-
beled data.
NBN estimates the more-likely class of each block
given only its calculated features. In step (d) of the al-
gorithm shown in Fig. 2, we start by determining the
next interesting feature to be computed in terms of its
cost and its relevance. Let F be the chosen feature
and H the other features computed in already itera-
tions. X = {F} ∪ H are used to infer the posterior
probability P
∗
= P(C = Char|X) using the following
rule:
P
∗
= P(C = Char|X )
=
1
Z
P(C = Char)
∏
x
j
∈X
P(x
j
|C = Char), (4)
with P(C = Char) the prior probability that B is
a character, and the evidence Z = p(X) = P(C =
Char)p(X|C = Char) + P(C = Non −Char)p(X |C =
Non −Char) is a scaling factor.
Thus, P
∗
will be high for characters and low for
non-characters. Then, P
∗
is compared with two deci-
sion pre-set thresholds T
1
and T
2
as follows: if P
∗
> T
2
then B corresponds to a character and if P
∗
< T
1
then B corresponds to a non-character, otherwise (if
T
1
< P
∗
< T
2
) the class of B can not be confirmed us-
ing only {F} ∪ H.
The undecided class corresponds to the confusing
difficult objects, which require additional verification.
Here, two scenarios are conceivable. Either there is
an error in the current segmentation. In this case, we
must revise the segmentation using devoted methods
(go from step (e) to step (c) in the algorithm shown
in Fig. 2). Either the calculated attributes are insuf-
ficient to confirm the final class. In this second case,
we suggest calculating and exploiting another feature
that has a high cost but may be more relevant such as
x
6
and x
7
(go from step (e) to step (d) of the algorithm
in Fig. 2).
We should note that the low-cost features
{x
1
,x
2
,x
3
,x
4
,x
5
} suffice to decide with confidence on
most blocks; additional re-segmentation or high-cost
features will still be demanded to deal with just a
small number of difficult objects. As a demonstra-
tion, we have designed and implemented a classifica-
tion algorithm that checks if the block B corresponds
to connected characters. This verification algorithm
is solicited in step (e) of the algorithm shown in Fig.
2.
3.2 Proposed Features
The proposed features x
1
,..., x
5
can be computed us-
ing simple formulas. Thus, in this section, we intro-
duce the 02 high-cost features x
6
(VGG16 score), and
x
7
(2D-DTW image matching distance).
3.3 VGG16 Score
VGG16 (Simonyan and Zisserman, 2014) is a simple-
fast-accurate CNN model for image classification. Its
architecture is depicted in Fig. 3. See the survey
(Shashirangana et al., 2020) to cover the proposed
deep learning methods, including the VGG variants
for license plate recognition.
In this work, the blocks’ images are resized
to 224x224 pixels, so they match exactly with the
VGG16 input shape. Blocks extracted from the train-
ing images are visually checked and labeled as ”char-
acter” or ”non-character” and then used to carry out a
transfer learning of the base model VGG16.
Improving the License Plate Character Segmentation Using Naïve Bayesian Network
63

Figure 2: Character segmentation algorithm.
Figure 3: VGG16 architecture.
3.4 Dynamic Time Warping
3.4.1 One-Dimensional Dynamic Time Warping
DTW (Berndt and J.Clifford, 1994) is a distance usu-
ally used for sequence matching and optimal align-
ment. The distance DTW (X,Y ) between two se-
quences X = (x
1
,x
2
,..., x
n
) and Y = (y
1
,y
2
,..., y
m
) can
be calculated by the construction of an n −by − m ma-
trix D of alignment, as shown in Fig. 4. The challenge
consists in finding the optimal path from cell (1,1)
to cell (n,m) having the minimal average cumulative
cost. Each cell (i, j) contributes cost(x
i
,y
j
) to the cu-
mulative cost. The cost function cost(x
i
,y
j
) may be
defined according to the application.
To perform computation efficiently, we can first
initialize D(1,1) to cost(1,1). Then, the optimal path
can be calculated using dynamic programming by em-
ploying the recursive equation shown in Eq. (5):
D(i, j) = Cost(i, j) + min



D(i − 1, j)
D(i, j − 1)
D(i − 1, j − 1)
(5)
The three previous stages, D(i − 1, j), D(i, j − 1),
and D(i − 1, j − 1), are assumed to be evaluated, and
their results are stored when computing D(i, j). No-
tice that the rest of the cells in the first column and
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
64

Figure 4: 1D-DTW example.
the first row have just one previous stage. In doing
so, DTW tolerates one-many and many-one match-
ing and allows thus to align sequences with differ-
ent lengths. Moreover, DTW can be speeded by lim-
iting the exploration to cells close to the diagonal
(1,1) ←→ (n, m).
In the example depicted in Fig. 4, the function
cost(x
i
,y
j
) = abs(x
i
,−y
j
) is considered. The optimal
alignment path, between sequences X and Y , is high-
lighted on the matching matrix, where its length is 14,
and its cumulative cost is 2.
3.4.2 Two-Dimensional Dynamic Time Warping
Indeed, extending 1D-DTW to 2D-DTW is so inter-
esting to match elastically images, which means that
the alignment should be invariant to image resizing
and geometrical transformations that preserve local
neighborhood features. In this section, we present
a concise introduction to 2D-DDW (Lei and Govin-
daraju, 2004), a variant of 2D-DTW for image match-
ing.
In a similar way, we can define D(i
1
, j
1
,i
2
, j
2
) as
the average cumulative cost matching between the
partial image (1 : i
1
,1 : j
1
) of image I
1
and the par-
tial image (1 : i
2
,1 : j
2
) of image I
2
as shown in Eq.
(6):
D(i
1
, j
1
,i
2
, j
2
) =
min
k:1..15
{Cost
k
(i
1
, j
1
,i
2
, j
2
)+
PreviousStage
k
}
(6)
Table 1 displays the possible values of the pre-
vious stages and their respective costs. 15 previous
stages are involved in the two-dimensional model of
Eq. (6) instead of 03 previous stages in the one-
dimensional model illustrated in Eq. (5). The cost
Cost
k
(i
1
, j
1
,i
2
, j
2
) is not the same for all stages as in
the 1D-DTW case, but it is defined depending on the
selected previous stage k. We have used the function
cost (Lei and Govindaraju, 2004) especially proposed
for image matching.
Table 1: Computing of Cost
k
(i
1
, j
1
,i
2
, j
2
) in the 2-DDW
recurrence rule.
k PreviousStage
k
Cost
k
(i
1
, j
1
,i
2
, j
2
)
1 D(i
1
− 1, j
1
− 1,i
2
− 1, j
2
− 1) R +C
2 D(i
1
− 1, j
1
− 1,i
2
− 1, j
2
) R +C
3 D(i
1
− 1, j
1
− 1,i
2
, j
2
− 1) R +C
4 D(i
1
− 1, j
1
− 1,i
2
, j
2
) R +C
5 D(i
1
− 1, j
1
,i
2
− 1, j
2
− 1) R +C
6 D(i
1
− 1, j
1
,i
2
− 1, j
2
) R
7 D(i
1
− 1, j
1
,i
2
, j
2
− 1) R +C
8 D(i
1
− 1, j
1
,i
2
, j
2
) R
9 D(i
1
, j1 − 1,i
2
− 1, j
2
− 1) R +C
10 D(i
1
, j1 − 1,i
2
− 1, j
2
) C
11 D(i
1
, j1 − 1,i
2
, j
2
− 1) C
12 D(i
1
, j
1
− 1,i
2
, j
2
) R +C
13 D(i
1
, j
1
,i
2
− 1, j
2
− 1) R +C
14 D(i
1
, j
1
,i
2
− 1, j
2
) R
15 D(i
1
, j
1
,i
2
, j
2
− 1) C
where R = 1D-DTW (I
1
(i
1
,:),I
2
(i
2
,:)) and
C = 1D-DTW (I
1
(:, j
1
),I
2
(:, j
2
))
Table 2: Segmentation results of the test images by the pro-
posed algorithm.
Predicted class
Real class Character Non-character Undecided
Character 2950 5 26
Non-character 4 10 5
Table 3: Segmentation results of the test images by the
YOLOv5 model.
Predicted class
Real class Character Non-character Undecided
Character 2963 9 9
Non-character 7 7 5
3.5 Experimental Results
A series of experiments are carried out on a dataset
including 850 vehicle images with Algerian license
plates, each measuring 600x840 pixels. 300 images
drawn randomly are used for the test; the rest (550)
are used to estimate the NBN and the VGG16 model
parameters.
The CS test results using our method are checked
visually, a summary is given in Table 2. The rate of
correct segmentation is (2950 +10)/(2950+5+26+
4 +10+ 5) = 98.67% which is accurate enough com-
pared with the quality of images.
We repeated the same tests on the same images us-
ing the YOLO-v5 model (Jocher, 2020). YOLOv5 is a
fast and accurate CNN model for detecting objects in
images, it is known as one of the leading solutions for
this task. For its use, we retrained it using the train-
ing sample containing 550 labeled images. The test
results obtained are displayed in Table 3. For the pre-
Improving the License Plate Character Segmentation Using Naïve Bayesian Network
65

diction, following the same approach as the proposed
solution, we retrieved from YOLO output the proba-
bility P(B = Char) that a detected block B is charac-
ter. Then, the block B is assigned to the undecided
class˝if P(B = Char) is between two thresholds T
1
,
T
2
(i.e. non-confirmation case).
In terms of accuracy, the results in Tables 2 and 3
are quite similar, with a slight advantage to YOLOv5.
Our method tends to struggle with separating con-
nected characters, often assigning them to the unde-
cided class. Conversely, YOLOv5 sometimes fails to
recognize new non-character objects and incorrectly
classifying them as characters or undecided.
Both algorithms are tested on the same workstation
(CPU: Intel i5 3.7GHz, RAM:32Go). In terms of
speed, the average time to process an image of a
plate is 32ms for the proposed method and 45ms
for YOLOv5. The proposed method is faster than
YOLOv5 on the CPU.
Table. 4 illustrates the segmentation and recogni-
tion (OCR) outputs of a set of localized plate exam-
ples using the proposed method. A deep visual anal-
ysis of these results has revealed the following facts:
1. the majority of errors of kind non-character pre-
dicted as Character corresponds to decorative or
publicity drawings often laid in particular places.
2. the majority of errors of kind Character predicted
as non-character corresponds to characters which
are excessively damaged or which are transcribed
in a font too different from that one of the used
characters prototypes.
3. The undecided objects usually correspond to
linked characters, damaged characters, partial
characters, or errors in the binarization step in the
pre-processing stage.
Linked characters and partial characters prob-
lems often have their origin at the level of the pre-
processing or the initial segmentation (steps (a) and
(b) in Fig. 2). Our solution has the advantage of trying
to classify the blocks into three classes instead of two.
The third class corresponds to confused objects (un-
decided) that need more examination. For this case,
we need to perform several checks to identify the type
of an undecided block (step (e) in Fig. 2). Then, de-
pending on that, our method can go back and revise
the binarization or segmentation (step (c) in Fig. 2).
In this step, we can split a block, merge two blocks, or
resize a block. Otherwise, it can involve supplemen-
tary features (step (d) in Fig. 2).
To illustrate, we suggest adding a two-linked char-
acters check module for the undecided objects. A
block B has a high potential to embody two charac-
ters or more when P(C = Char|x
1
,x
3
,x
4
,x
6
) is high,
P(C = Char|x
6
,x
7
) is low, and x
2
is high (i.e. large
blocks). In this case, the block is split into two ad-
jacent blocks B
le f t
and B
right
so that B
le f t
has a high
probability to be a unique character. The algorithm
shown in Table 5, analyses the content of B and might
split it using the robust 2D-DTW matching distance.
It divides B at the horizontal position i if the distance
2D-DTW(B(2:i,:), character templates) is very low.
Obviously, other checks can be easily integrated.
In our experiment, the two-linked check module
has allowed us to split 18/29 connected characters
successfully.
3.6 Complexity Analysis
The computation time of the 2-DWW score is signifi-
cantly higher compared to the rest of the steps in our
algorithm. In this section, we estimate the complexity
of the proposed 2-DDW implementation.
Suppose both images I
1
and I
2
are N-by-N. The
complexity of constructing the matching matrix by
04 nested loops is α = O(N
4
) iterations. We need
to determine the cost value of the 15 previous stages
at each iteration. It is worth noting that a stage may
be the previous of many stages; thus, speeding the
computation is possible by saving and reusing inter-
mediate cost values.
According to (Lei and Govindaraju, 2004), the
evaluation of Cost
k
(i
1
, j
1
,i
2
, j
2
) involves the evalua-
tion of 1D-DTW between rows I
1
(i
1
,:) and I
2
(i
2
,:) or
between columns I
1
(:, j
1
) and I
2
(:, j
2
). In the worst
case, 1D-DTW is computed once at most between
each row pair and between each column pair of im-
ages I
1
and I
2
. The complexity of computing all
stages’ costs is β = 2O(N
2
)×O(1D-DTW ) = O(N
4
).
The entire complexity is then α + β = O(N
4
). One of
our main contributions consists in decreasing the time
complexity from O(N
6
) (Lei and Govindaraju, 2004)
to O(N
4
) using the dynamic programming principle.
3.7 Future Improvement
In this work, we have illustrated a robust segmenta-
tion framework that quantifies the segmentation qual-
ity by assessing the probability that a block corre-
sponds to a character and reducing the processing
time. These two aspects are interesting to develop
new ALPR systems analyzing video sequences where
cars move continuously and the image acquisition
conditions vary. This characteristic may help com-
bine partial results of images covering identical vehi-
cles efficiently and create real-time solutions.
We suggested predicting the class of each block
among the three following labels: ”confirmed as a
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
66

Table 4: Examples of character segmentation and recognition.
Localized plate Pre-processing Segmentation Recognition
0120230416 success
0429319909 success
0930010716 success
3564779916 success
416431078 failed
1134610516 success
8516131861 failed
0994130016 success
3346810816 success
Table 5: Algorithm that checks and splits two-linked char-
acters.
Input: B, the block to be checked and split.
Output: B
le ft
,B
right
// B may be split into B
le ft
and B
right
K ← width of B;
for i = 1 to K do
D(i) ← min
j:0..9
2-DDW (B(1 : i, :),λ
j
);
//λ
j
: template of digit
′
j
′
t ← index of (min(D));
If D(t) < threshold, then % split B
B
le ft
← B(1 : t,:);
B
right
← B(t + 1 : K,:);
character”, ”confirmed as a non-character”, and ”un-
decided”. CS errors may be due to pre-processing,
NBN classification, or undecided object processing
steps deficiency. In future works, we recommend the
following directions:
1. defining new cost-effective features to minimize
the number of undecided objects. For example,
global features on the order and distances between
successive characters appear to be beneficial.
2. finding the best order of features for each block.
For example, we can create a cascade of the low-
cost classifiers as shown in the CASACRO algo-
rithm (Hanczar and Bar-Hen, 2021).
3. examining various types of undecided blocks and
suggesting suitable processing, including the pos-
sibility of re-segmentation.
4 CONCLUSION
In this study, we have developed a robust character
segmentation method for vehicle license plates. Ini-
tially, we employ a standard CS technique such as
CCA, projection, or CNN to identify the potential
characters. Subsequently, we utilize the NBN clas-
sifier to predict the label of each block, categoriz-
ing them as ”Character”, ”Non-Character”, or ”Unde-
cided”. In addition, the proposed algorithm provides
the probability that a block corresponds to a charac-
ter, which is valuable in various applications. How-
ever, different features can have varying impacts on
the classification process, and their computation in-
curs different time costs. Therefore, the most compu-
tationally expensive features should be involved only
for a few for challenging or ambiguous blocks.
We emphasize that, in the problem we’re address-
ing, only a few low-cost features are computed for
most blocks, and almost all the features are approx-
imately independent conditionally on the class. Our
primary objective is to strike a balance between accu-
racy and speed.
To meet these criteria, we recommend using the
NBN classifier as it aligns perfectly with these as-
sumptions and requirements. In this study, we
have also introduced a faster variant of the 2-DDW
template matching algorithm (Lei and Govindaraju,
2004), significantly reducing complexity to O(N
4
).
Improving the License Plate Character Segmentation Using Naïve Bayesian Network
67

State-of-the-art CNN models, like YOLOv5, offer
several advantages, including ease of implementation,
high recognition rates, and running faster on GPU.
They are challenging to surpass in these aspects. Nev-
ertheless, the proposed method presents distinct ad-
vantages. It operates as a white box, where its mod-
ules and steps can be explained. Additionally, it runs
efficiently on the CPU, and it can be enhanced in the
future in various aspects.
We have introduced a general approach to CS. Our
approach allows for the utilization and integration of
segmentation methods published in the literature. The
core principles of our proposal are as follows: i) We
assess the probability that a block is either a character,
not a character, or uncertain base on the pre-computed
features. To accomplish this, we have chosen to em-
ploy Bayesian networks due to their capability to han-
dle missing features and assess class probabilities in
a formal manner. ii) The introduction of the ”unde-
cided block” class is significant, as it simplifies the
process of either requesting additional features or ad-
justing the segmentation. iii) We can incorporate prior
information regarding the structure of license plates.
Thus, it is less sensitive to the over-learning problem
often encountered for deep neural network models.
Overall, obtained partial results are promising.
For instance, further improvements by providing new
features and comparisons with state-of-the-art meth-
ods using international public datasets are recom-
mended to highlight the proposed approach.
REFERENCES
Berndt, D. and J.Clifford (1994). Using dynamic time warp-
ing to find patterns in time series. In Working Notes
of the Knowledge Discovery in Databases Workshop,
pages 359–370.
Du, S., Ibrahim, M., Shehata, M., and Badawy, W. (2013).
Automatic license plate recognition (alpr): A state-of-
the-art review. IEEE Trans. on Circuits and Systems
for Video Technology, 23:311–325.
Hanczar, B. and Bar-Hen, A. (2021). Cascaro: Cascade of
classifiers for minimizing the cost of prediction. Pat-
tern Recognition Letters, 149:37–43.
He, X., Zheng, L., Wu, Q., W.Jea, Samali, B., and
Palaniswami, M. (2008). Segmentation of characters
on car license plates. In Proc.IEEE 10th Workshop on
Multimedia Signal Processing, pages 399–402.
Jagannathan, J., Sherajdheen, A., vijay deepak, R. M.,
and Krishnann, N. (2013). License plate character
segmentation using horizontal and vertical projection
with dynamic thresholding. In IEEE Inter. Conf. on
Emerging Trends in Computing, Communication and
Nanotechnology, pages 700–705.
Jocher, G. (2020). Yolov5 by ultralytics.
Lei, H. and Govindaraju, V. (2004). Direct image matching
by dynamic warping. In Computer Vision and Pattern
Recognition Workshop.
Marcel, B. (1987). Geometry I. Berlin: Springer.
Mittal, A., Kassim, A., and Tan, T. (2007). Bayesian Net-
work Technologies: Applications and Graphical Mod-
els, chapter chapter 1, pages 1–12. IGI Publishing,
Hershey New York.
Montazzolli, S. and Jung, C. (2017). Real-time brazilian li-
cense plate detection and recognition using deep con-
volutional neural networks. In 2017 30th SIBGRAPI
conference on graphics, patterns and images (SIB-
GRAPI), pages 55–62. IEEE.
Mufti, N. and Shah, S. A. A. (2021). Automatic number
plate recognition: A detailed survey of relevant algo-
rithms. Sensors, 21(9):3028.
Pan, M.-S., Yan, J.-B., and Xiao, Z.-H. (2008). Vehicle
license plate character segmentation. Inter.Journal of
Automation and Computing, 05(04):425–432.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 779–
788.
Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster
r-cnn: Towards real-time object detection with region
proposal networks. Advances in neural information
processing systems, 28.
Sauvola, J. and PietikaKinen, M. (2000). Adaptive docu-
ment image binarization. Patt. Recog., 33:225–236.
Shashirangana, J., Padmasiri, H., Meedeniya, D., and Per-
era, C. (2020). Automated license plate recognition:
a survey on methods and techniques. IEEE Access,
9:11203–11225.
Simonyan, K. and Zisserman, A. (2014). Very deep con-
volutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556.
Yoon, Y., Ban, K.-D., Yoon, H., and Kim, J. (2011).
Blob extraction based character segmentation method
for automatic license plate recognition system. In
Proc.IEEE International Conference on Systems,
Man, and Cybernetics, pages 2192–2196, Korea.
ICINCO 2023 - 20th International Conference on Informatics in Control, Automation and Robotics
68