LEARNING SPLINE MODELS WITH THE EM ALGORITHM FOR SHAPE RECOGNITION

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.14, No.6, November 2023
DOI:10.5121/ijaia.2023.14604 37
LEARNING SPLINE MODELS WITH THE EM
ALGORITHM FOR SHAPE RECOGNITION
Abdullah A. Al-Shaher, Yusef S. AlKhawari
Department of Computer Science and Information Systems, College of Business Studies,
Public Authority for Applied Education and Training, Kuwait
ABSTRACT
This paper demonstrates how cubic Spline (B-Spline) models can be used to recognize 2-dimension non-
rigid handwritten isolated characters. Each handwritten character is represented by a set of non-
overlapping uniformly distributed landmarks. The Spline models are constructed by utilizing cubic order of
polynomial to model the shapes under study. The approach is a two-stage process. The first stage is
learning, we construct a mixture of spline class parameters to capture the variations in spline coefficients
using the apparatus Expectation Maximization algorithm. The second stage is recognition, here we use the
Fréchet distance to compute the variations between the spline models and test spline shape for recognition.
We test the approach on a set of handwritten Arabic letters.
KEYWORDS
Spline Models, Cubic order of Spline Curves, handwritten Arabic Characters, Shape Recognition,
Recognition by Alignment. Expectation Maximisation Algorithm, Unsupervised Learning, Fréchet distance.
1. INTRODUCTION
Shape matching has been the focus of many researchers for the past seven decades [1] and
attracted many scholars in the field of pattern recognition [2], artificial intelligence [3], signal
processing [4], image analysis [5], and computer vision [6]. The difficulties arise when the shape
under study exhibits a high degree in shape variation: as in handwritten characters [7], digits [8],
face detection [9], and gesture authentication [10]. For a single data, shape variation is limited and
cannot be captured ultimately because single data does not provide sufficient information and
knowledge about the data; therefore, multiple existence of data provide better understanding of
shape analysis and manifested by mixture models [11]. Because of the existence of multivariate
data under study, there is always the requirement to estimate the parameters that describe the data
that is encapsulated within a mixture of shapes. Extensive efforts have been invested in analyzing
alphabet of English [12], Latin [13], Chinese [14], Hebrew [15], and Hindi [16]; however, very
little work has been done in analyzing handwritten Arabic characters and least considerable
results achieved [17]. The subject of investigating Arabic letters shape analysis is complicated.
Such difficulties arise from the nature structure of Arabic alphabet and words. Arabic script is
cursive, continues, and similar without diacritics. For example: the word ‫م‬ ‫عل‬ has different
meaning when secondaries are attached to the letters.
The literature demonstrates many statistical and structural approaches with various algorithms to
model shape variations using supervised and unsupervised learning [18] algorithms. In precise,

38
comes the powerful Expectation Maximization algorithm (EM) of Dempster [19] that has been
used widely for such cases. The EM algorithm revolves around two step procedures. The
expectation E step revolves around estimating the parameters of a log likelihood function and
pass it to the Maximization M step. In a maximization (M) step, the algorithm computes
parameters maximizing the expected log-likelihood found on the E step, the process is iterative
one until all parameters come to stability. For instance, Jojic and Frey [20] have used the EM
algorithm to fit mixture models to the appearance manifolds for faces. Bishop and Winn [21]
have used a mixture of principal components analyzers to learn and synthesize variations in facial
appearance. Vasconcelos and Lippman [22] have used the EM algorithm to learn queries for
content-based image retrieval. Finally, several authors have used the EM algorithm to track
multiple moving objects [23]. Revov et al. [24] has developed a generative model which can be
used for handwritten character recognition. Their method employs the EM algorithm to model the
distribution of sample points.
Spline models have been used widely in the signal processing community [25]. Spline models can
describe more complex shapes that contain circles, closed contours, ellipses, and curved shapes
and combine this description in mathematical and statistical framework. For example, Splines are
used to model Functional data analysis [26], Computer Aided Design [27], automobiles [28],
smoothed surfaces [29], curly shapes [30], and noisy materials such as clothing design [31].
Safraz [32] has investigated the issue of Arabic shape recognition by utilizing edge detection of
machine printed letters to bring the shape under landmarks representation, then applying cubic
polynomials on them. In deepest investigation, Safraz [33] addressed the subject of extracting the
curve estimation by finding the optimal set of coefficients of Bezier function and dividing the
curve into multiple segments ending with a smooth curve fitting.
In this research paper, we demonstrate how Spline models are used to recognize 2D handwritten
shapes by fitting 3rd order of polynomial function to a set of landmarks points extracted for the
shape under analysis. We then, train such Spline models to capture the optimal characteristics of
the shapes in the training sets of shapes. Handwritten isolated Arabic characters are used and
tested in this investigation.
2. SHAPE MODELLING
A spline curve is a mathematical representation to construct an interface that allows sequence of
points to design and control the shape under study.
The parametric form of shape coordinates is called control points. The core concept of pattern
recognition methods and approaches aim at constructing a classifier for a set of training patterns.
Such a classifier determines which shape-class the input pattern belongs to; therefore, is known as
object training or shape learning [34][35]. In real world applications such as object recognition, a
certain previous knowledge is required beforehand. The utilization of such knowledge into a
training set is the key element that will allow to increase of performance of shape classification.
Following the concept of constructing a training set of patterns undergo classification. The
landmark patterns are collected as the object in question undergoes representative changes in
shape. To be more formal, each landmark pattern consists of L labelled points. Suppose that there
are T landmark patterns. The tth training pattern is represented using the long vector of landmark
co-ordinates:
(11)

39
where the subscripts of the co-ordinates are the landmark labels. A spline curve is a curve that
passes through these tth training pattern co-ordinates called interpolating curves. To establish
interface, we approach the polynomial in the form of function of landmarks:
(12) +
…..
The degree of the polynomial corresponds with the highest coefficient that is nonzero. The aim of
using cubic spline is to provide a smooth curve that passes through the chosen landmarks for
shape construction as well as support inflection points. Equation (12) can be revised in the form
of cubic polynomials.
(13)
Equation (13) can be rewritten in the form of solving linear equations to solve for the spline
coefficients obtained by the curve segment of 4 coordinates. The new mathematical
formulation to solve for required coefficients, we present a matrix of linear equations.
(14)
Where is the matrix of linear equations
(15)
The solution is
(16)
Similar procedure is applied to the remaining coordinate segments. The resulting set of spline
coefficients presented by
(17)
And in unification form of the set spline coefficients to be
) (18)
Suppose that each shape class have T training spline mean coefficients . The mean vector
of is represented by

40
(19)
The covariance matrix is then constructed.
(20)
3. LEARNING SPLINE MODEL
In machine learning approaches [21], supervised learning depends on data labelling which the
outcomes models are mapping algorithms between data and its counterparts. On the other hand,
unsupervised learning data patterns assumed from the unlabeled input data, hence, the goal of
unsupervised learning is to find the structure and variations from the input data. In another format,
Unsupervised learning does not need any supervision architecture. Instead, it finds data variations
and variable modes from the data by itself. To reproduce more complex variations in shape either
a non-linear deformation or a series of local piecewise linear deformations must be employed. In
this paper we adopt an approach based on mixtures of spline model distributions. Our reasons for
adopting this approach are twofold. First, we would like to be able to model more complex
deformations by using multiple modes of shape deformation. This arises when a set of training
patterns contains examples from different shape classes. The second reason is that shape
variations cannot be captured in a single mode of shape deformation, hence a mixture is
constructed. We commence by assuming that the individual examples in the training set are
conditionally independent of one another. The following approach is based on fitting a Gaussian
mixture model to the set of training examples of spline coefficients. We further assume that the
training data can be represented by a set of spline-classes Ω.
Each spline coefficient model ꞷ has its own mean coefficients and covariance matrix .
With these ingredients we establish the likelihood function for the set of the curve patterns in
(21)
Where the term is the probability for drawing curve pattern αt from the curve-
class ω. Associating the above likelihood function with the Expectation Maximization algorithm,
the likelihood function can be written to be iterative process of two steps. The process revolves
around estimating the expected log-likelihood function iteratively in
(22)
Where the quantity and are the estimated mean curve vector and variance
covariance matrix both at iteration (n) of the algorithm. The quantity is the
a posteriori probability that the training pattern spline coefficients belong to the spline-class
and

41
ω at iteration n of the algorithm. The term is the probability of
distribution of spline-pattern αt belonging to spline-class ω at iteration ( n + 1 ) of the algorithm;
thus, the probability density to associate curve- patterns for ( t = 1 … T ) to class spline-class
ω are estimated by the updated construction of the mean-vector , and
covariance matrix at iteration n+1 of the algorithm. According to the EM algorithm, it
revolves around estimation of the expected log-likelihood function within two iterative processes.
In the M or maximization step of the algorithm, our aim is to maximize the curve
mean-vector , and covariance matrix , while, in the E or expectation step, the aim
is to estimate the distribution of spline-patterns at iteration n along with the mixing proportion
parameters for curve-class ω.
In the E, or Expectation step of the algorithm, the a posteriori spline-class probability is updated
by applying the Bayes factorization rule to the spline-class distribution density at iteration ( n+1 ).
The new estimate is computed by
(23)
Where the revised spline-class ω mixing proportions at iteration ( n + 1 )is computed in
(24)
With that at hand, the distributed spline-pattern t to the class-spline ω is Gaussian distribution
and is classified according to
(29)
In the M, or Maximization step, our aim is to maximize the curve-class ω parameters. The
updated spline mean-vector estimate is computed by
(30)

42
And the new estimate of the spline-class covariance matrix is weighted by
(31)
Both E, and M steps are iteratively converged, the outcome of the learning stage is a set of spline-
class ω parameters such as , hence the complete set of all spline-class Ω are
computed and ready to be used for recognition.
4.SPLINE MATCHING
With the ingredients produced by the previous learning stage, the Spline model parameters
belongs to the set of spline-classes Ω can be used for the purpose of recognition.
To establish a mathematical setting, each control point in the test spline is mapped to its
counterparts in the spline model. Hence, we use the Fréchet distance [36].
Fréchet distance evaluates the similarity between two spline curves. Let's denote is the spline
mean and data spline data curve. Fréchet distance is defined as the minimum cordlength
sufficient to map two control points in both splines. The mapping continues until reaching the end
of the spline curves. The distance is then the end of point traced forward along and one
traveling forward along , although the rate of travel for either point may not
necessarily be uniform.
4.1. Spline Projection
Assume that the class-mean spline and the data spline curve are projected into 2
dimension coordinate space by estimating the y axes over
the spline coefficients. the resulting vector of shape representation for the class-mean spline
denoted by
(32)
And test-data spline projection is denoted by
(33)
4.2. Fréchet Matrix
The matrix of Fréchet distances is then computed for each pair of the control points of the class-
mean spline and the test-data spline in

43
( 34)
where
(35)
is the Euclidean distance between any pair of coordinates space. The matrix constructed is square
non-symmetric matrix. We consider the diagonal cells of the matrix that explain the distance
between control points pair for both objects under study, we choose the maximum value as a top
limit for the final distance. The algorithm proceeds by iterating through the diagonal left to right,
and for each column, calculates the distances and stores them if they are smaller than the
reference maximum. The same process applies to rows.
(36)
These values are compared with the other class-mean spline in the set , the class-mean spline
that has the minimum values is selected data spline is then grouped for that class.
5. EXPERIMENTS
We have evaluated our approach with sets of isolated Arabic handwritten characters. Here, we
have used 22 shape-classes for different writers, each with 40 training patterns. In total, we have
tested the approach with 880 handwritten Arabic character shape patterns for testing and 1760
patterns for testing our approach. Figure 1 shows some training patterns used in this research
paper. Figure 2 shows single shapes and their landmarks representation of being uniformly
distributed along the skeleton of the shape under investigation.
Figure 1: Sample handwritten character training sets.

44
Figure 2: Training patterns and extracted landmarks.
In figure 3, sample cubic spline coefficients resulted from the training stage for spline-class ꞷ.
Here, each raw represents the set of coefficients for the handwritten character ( Seen ) and for
each segment of the shape.
Figure 3: Spline coefficients for sample class-shape
In figure 4, illustrates the result from the training stage. First raw shows sample training sets, raw
2 shows the training stage Expectation Maximization initialization shape, raw 3, shows the
projection of the spline training coefficients producing the landmarks, and raw 4 show the spline-
mean for class shape representation. Figure 5 shows the convergence rate as a function per

45
iteration number. The graph shows how associated distributed probabilities for the set of spline-
classes Ω converged in a few iterations.
Figure 4: a) sample sets b) EM initialization c) Mean Shape landmarks
d) Spline Mean Shape
Figure 5: Convergence Rate as a function per iteration no.

46
Figure 6 shows the Fréchet distance table between the spline-mean shape and a testing shape
pattern.
The graph shows the similarity between the spline curves which considers the location and
ordering landmarks to the sample testing pattern landmarks. It is apparent from the table that the
similarity is shown in the diagonal elements highlighted by green color. The diagonal distance
between the landmarks for both the spline-mean shape and it's corresponding in the testing
pattern. The similarity excluded by the minimum distance.
Figure 6: Fréchet distance
To take the investigation further, we demonstrate how the approach behaves under the presence of
noise. In figure 7, we show how recognition rate is achieved when point position displacement
error is applied. Test shape coordinates are being moved away from their original position. The
figure shows the recognition rate fails to recognize shapes to their correct class ω in a few
iterations and it fails completely when coordinates are moved away, yet, increasing variance
significantly.
Figure 7: Recognition rate as a function per iteration no with point position error.

47
Table 1 shows recognition rates per spline-classes ω. Row 1 represents the shape name in Arabic
alphabet. Row 2 represents the character shape. Row 3 lists the size of the testing patterns. Rows
4,5 show the correct and wrong registration of shapes to their spline-class ω.
Row 6 shows the recognition rate per spline-class ω. In total, we have achieved 96% recognition
rate for such an approach.
Shape Name
Sample Shape Test Size Correct False
Recognition Rate
Ain_1 80 78 2 97.5%
Baa 80 78 2 97.5%
Dal 80 79 1 98.8%
Faa 80 73 7 91.3%
Haa_1 80 79 1 98.8%
Ttah 80 76 4 95%
Hhah_1 80 71 9 88.8%
Ain_2 80 79 1 98.8%
Meem 80 76 4 95%
Seen 80 74 6 92.5%
Yaa 80 77 3 96.3%

48
Haa_2 80 78 2 97.5%
Waw 80 76 4 95%
Ain_2 80 79 1 98.8%
Hhah_2 80 74 6 92.5%
Kaf_1 80 79 1 98.8%
Lam_1 80 80 0 100%
Lam_2 80 73 7 91.3%
Raa 80 79 1 98.8%
Ssad 80 73 7 91.3%
Kaf_2 80 78 2 97.5%
Noon 80 79 1 98.8%
Total 22 classes 1760 1688 72 96%
6. CONCLUSION
In this research paper, we have shown how cubic spline 3rd degree polynomials can be used for
recognizing handwritten Arabic characters. Each shape under testing is represented by a set of
uniformly distributed landmarks along the skeleton. A cubic piecewise B-Spline is implemented
on the shape to extract the coefficients needed for training set. A mixture of spline coefficients
has been trained in a two-level hierarchy Expectation Maximization setting to capture the modes
of spline parameters describing the pattern-sets. Recognition is then applied on a different set of
testing patterns using the Fréchet distance mechanism. We have shown a remarkable recognition
rate of 96%.

49
7. FUTURE WORK
In this research, there are some shortcomings to the approach. The first shortcoming is proving
cubic b-spline 3rd order of polynomials is best representation for handwritten Arabic characters.
The second is, the extracted parameters from the training stage have not been utilized for the
purpose of recognition stage encapsulated within a statistical framework: for example, spline
coefficients are not utilized for spline alignment. Secondly, there is a need to provide a
hierarchical framework for spline alignment.
REFERENCES
[1] J. Wood, "Invariant pattern recognition: A review," Pattern Recognition, vol. 29, no. 1, 1996, 1-17.
[2] Anil, K. Jain,Robert P.W. Duin, and Jianchang Mao: “Statistical Pattern Recognition: A Review”,
IEEE Pattern Analysis and Machine Intelligence, vol 22, No. 1, PP 4-37, 2000.
[3] P.F. Baldi and K. Hornik, "Learning in linear neural networks: A survey," IEEE Transactions on
Neural Networks, vol. 6, no. 4, 1995, 837-858.
[4] T.Y. Kong and A. Rosenfeld, "Digital topology: introduction and survey," Computer Vision,
Graphics, and Image Processing, vol. 48, no. 3, pp. 357-393, 1989.
[5] T.R. Reed and J.M.H. Dubuf, "A review of recent texture segmentation and feature extraction
techniques," CVGIP - Image Understanding, vol. 57, no. 3, pp. 359-372, 1993.
[6] S. Sarkar and K.L. Boyer, "Perceptual organization in computer vision - a review and a proposal for a
classificatory structure", IEEE Transactions on Systems Man and Cybernetics, vol. 23, no. 2, 1993,
382-399.
[7] Lorigo L. M., Govindaraju V., “Offline Arabic Handwriting Recognition: A Survey”, IEEE Trans
Pattern Anal Mach Intelligence, Vol: 28(5): PP 712-24, 2006.
[8] Ishani Patel, Virag Jagtap, Ompriya Kale, “A Survey on Feature Extraction Methods for Handwritten
Digits Recognition”, International Journal of Computer Applications, Vol 107, No 12, PP 1117,
2014.
[9] Muhammad Sharif, Farah Naz, Mussarat Yasmin, Muhammad Alyas Shahid, Amjad Rehman, “Face
Recognition: A Survey”, Journal of Engineering Science and Technology Review, Vol 10, No 2, PP
166-177, 2017.
[10] Rafiqul Zaman Khan, Noor Adnan Ibraheem, “HAND GESTURE RECOGNITION: A
LITERATURE REVIEW “,International Journal of Artificial Intelligence & Applications, Vol.3,
No.4, pp 161-174, 2012.
[11] Bruce George Lindsay,Mary L. Lesperance, “A review of semiparametric mixture models”, Journal of
Statistical Planning and Inference, Vol: 47 , No: 1 PP 29-39 , 1995.
[12] Patel, Shital P. Thakkar, "Handwritten Character Recognition in English: A Survey", International
Journal of Advanced Research in Computer and Communication Engineering, Vol. 4, Issue 2, 2015.
[13] Edyta LUKASIK, Malgorzata CHARYTANOWICZ, Marek MILOSZ, Michail TOKOVAROV,
Monika KACZOROWSKA, Dariusz CZERWINSKI, Tomasz ZIENTARSKI, "Recognition of
handwritten Latin characters with diacritics using CNN", ARTIFICIAL AND COMPUTATIONAL
INTELLIGENCE, Vol. 69. No 1, 2021.
[14] Dai Ruwei,Liu Chenglin,Xiao Baihua, "Chinese character recognition: history, status and prospects",
Frontiers of Computer Science in Chinavolume 1, pages126–136 (2007).
[15] Tabita L. Tobing, Sule Y. Yayilgan, Sony George, " Isolated Handwritten Character Recognition of
Ancient Hebrew Manuscripts ", SOCIETY FOR IMAGING SCIENCE AND TECHNOLOGY, PP
35-39, 2022.
[16] Madhuri Yadav,Ravindra Kumar, Purwar,Mamta Mittal, "Handwritten Hindi character recognition: a
review", IET Image Processing, Volume12,Issue11, November 2018, Pages 1919-1933.
[17] Pervez Ahmed and Yousef Al-Ohali, "Arabic Character Recognition: Progress and Challenges", J.
King Saud Univ., Vol. 12, Comp. & Info. Sci., pp. 85-116,2000.
[18] Christopher M. Bishop, “Neural Networks for Pattern Recognition”, Clarendon Press, 1995, ISSN
0198538642.
[19] A. Dempster, N. Laird, D. Rubin, "Maximum likelihood from incomplete data via the em algorithm",
J. Roy. Statist. Soc. Ser. 39 (1977) 1–38

50
[20] J. Brendan, N. Jojic, Estimating mixture models of images and inferring spatial transformations using
the em algorithm, IEEE Computer Vision Pattern Recognition 2 (1999) 416–422.
[21] C. Bishop, J. Winn, Non-linear Bayesian image modelling, Proceedings of Sixth European
[22] N. Vasconcelos, A. Lippman, A probabilistic architecture for content-based image retrieval,
Proceedings of International Conference on Computer Vision and Pattern Recognition, 2000, pp.
216–221.
[23] B. North, A. Blake, Using expectation-maximization to learn dynamical models from visual data,
Image Vision Computing. 17 (8) (1999) 611–616.
[24] M. Revow, C. Williams, G.E. Hinton, Using generative models for handwritten digit recognition,
IEEE Trans. Pattern Anal. Mach. Intell. 20 (2) (1996) 592–606
[25] M. Svoboda,L. Matiu-Iovan,P. Andea, B-spline interpolation technique for digital signal processing ,
International Conference on Information and Digital Technologies, 2015
[26] A.M. Aguilera, M.C. Aguilera-Morillo, Comparative study of different B-spline approaches for
functional data. Mathematical and Computer Modelling 58 (2013) 1568–1579.
[27] E. Dimas, D. Briassoulis, 3D geometric modeling based on NURBS: a review, Journal of Advances in
Engineering Software, (1999)741-751.
[28] A. Galvez, A. Iglesias, A new iterative mutually coupled hybrid GA-PSO approach for curve fitting in
manufacturing, Journal of Applied Soft Computing, (2012),
[29] I. Zeid, Mastering CAD/CAM, McGraw Hill International Edition, New York, 2005.
[30] M. Sevaux, Y. Mineur, A curve-fitting genetic algorithm for a styling application, European Journal
of Operational Research, 179 (2007) 895-905.
[31] R. Liutkevicius, A. Davidsonas, Surface Reconstruction from Partially Structured Noisy Cloud of
Points Using B-Splines, Journal of Information Technology and Control, 42 (2013).
[32] Muhammed Safraz, Mutaza Khan, “Automatic outline capture of Arabic fonts”, Journal of
information Sciences, Vol 140, No. 3-4, PP 269-281,2002
[33] Muhammad Safraz, “Optimal Curve Fitting to Digital Data”, Journal of WSCG, Vol 11, No. 1-3,
2003.
[34] AlShaher Abdullah, Hancock Edwin, Learning mixtures of Point Distribution models with the EM
algorithm. Pattern Recognition, 36(2003) 2805-2818.
[35] Christopher M. Bishop, Pattern Recognition and Machine Learning, Springer Science, and Business
Media, 2006.
[36] Thomas Heiter and Heikki Manila. Computing Discrete Fréchet Distance. Technical Report CD-TR
94/64, Christian Doppler Laboratory for Expert Systems, Technical University Vienna, Austria, 1994.
AUTHORS
ABDULLAH A. AL-SHAHER is currently enrolled as associate professor faculty
member in the Computer Science and Information Systems Department at the college of
Business Studies in the Public Authority for Applied Education and Training, Kuwait. He
received his Ph.D. in Computer Science from the University of York, York, United
Kingdom in 2005. In 1995 he was awarded master's degree (honour list) in Computer
Science from University of Denver, Denver, Colorado, USA, and a bachelor's degree from
HustonTillotson University in 1988 Austin, Texas, USA. His main interest is Shape
Recognition, Statistical Patten Recognition, Computer Vision, and Artificial Intelligence.
YOUSEF S. ALKHAWARI is currently enrolled as a lecturer faculty member in the
Computer Science and Information Systems Department at the College of Business Studies
in the Public Authority for Applied Education and Training (PAAET), Kuwait. He
received his master's degree in computer science from California State University of
Sacramento (CSUS), Sacramento, California, USA in 1995. In 1991 he earned his
bachelor's degree in computer engineering from University of the Pacific (UOP), Stockton,
California, USA. His main interest is Knowledge Based Systems and Artificial Intelligence
and machine learning.

LEARNING SPLINE MODELS WITH THE EM ALGORITHM FOR SHAPE RECOGNITION

More Related Content

Similar to LEARNING SPLINE MODELS WITH THE EM ALGORITHM FOR SHAPE RECOGNITION (20)

More from gerogepatton (20)

Recently uploaded (20)

LEARNING SPLINE MODELS WITH THE EM ALGORITHM FOR SHAPE RECOGNITION