Lanna Dharma Printed Character Recognition using k-Nearest
Neighbor and Conditional Random Fields
Chutima Chueaphun
1
, Atcharin Klomsae
1
, Sanparith Marukatat
2
and Jeerayut Chaijaruwanich
1
1
Department of Computer Science, Faculty of Science, Chiang Mai University, Chiang Mai, 50200, Thailand
2
National Electronics and Computer Technology Center, Pathumthani, 12120, Thailand
Keywords: Lanna Dharma, Character Recognition, k-Nearest Neighbor, Conditional Random Fields.
Abstract: For centuries, in the North of Thailand, many books of Lanna Dharma characters had been printed. These
books are the important sources of the knowledge of ancient Lanna wisdom. At present, the books are found
old and damaged. Most of characters are rough and not clear according to its early printing technology at
that time. Moreover, some sets of characters are relatively very similar which cause the difficulty to
recognize them. This paper proposes a Lanna Dharma printed character recognition technique using k-
Nearest Neighbor and Conditional Random Fields. The accuracy of recognition rate is about 82.61 percent.
1 INTRODUCTION
Currently, there are many optical character
recognition (OCR) researches which allow
conversion of the text in scanned images into the
machine-encoded text. OCR systems are available in
many languages such as English, Japanese, Chinese,
Arabic, Thai, etc. However, there are not yet any for
Lanna Dharma character.
Hundreds of years ago, Lanna language was
widely used in the northern part of Thailand during
the time of the Lanna kingdom, which was founded
in 1259. The Lanna Dharma character is a
descendant of the old Mon character like Lao and
Burmese characters. Since 1892, the typed Lanna
Dharma character was first printed as books
including history, medicine, literature and
Buddhism. The Lanna Dharma books are the
important sources of Lanna regional knowledge.
However, after the invading of Ayutthaya kingdom
from the central of Thailand, the Central Thai
language became the official language learned in
school. Now a day, the Lanna Dharma character has
almost been forgotten. There are now only few
people, usually old ones, who can read it. In
addition, most of the Lanna Dharma books are lost
and destroyed. Therefore, Lanna Dharma printed
character recognition will help to preserve the
ancient Lanna knowledge. Furthermore, the
knowledge can be delivered to general public with
electronically retrievable.
The Lanna Dharma writing system has no white-
space between each word, but there is a white-space
at the end of a clause or sentence. Lanna Dharma
word consists of consonants, vowels, and tones at
different levels. Specifically, many consonants have
their alternate form when they directly follow other
consonant. We distinguish the types of characters for
this paper into eight different types as shown in
Table 1.
The writing example of Lanna Dharma word is
shown in Figure 1. It can be written by organizing
the consonants or the middle vowels in level 1; the
upper vowels are in level 2; the tones are in level 2
or 3; the final consonants are in level 4 and the lower
vowels are in level 4 or 5.
In this paper, we mainly focus on the solution to
the problem of distinguish character belonging to
confusion sets. Indeed, Lanna Dharma printed
character have relatively similar patterns which
cause the recognition error. Figure 2 shows the
example of confusion sets. We propose the use of k-
Nearest Neighbor (k-NN) to firstly classify the class
of character images. Then, sequence of character
classes which is the output of k-NN is reclassified
again by Conditional Random Fields (CRFs). CRFs
are the conditional undirected graphical models,
which model the conditional probabilities of
character sequence. Therefore, it is expected to
resolve the problem of k-NN for confusion sets and
improve the final character recognition.
169
Chueaphun C., Klomsae A., Marukatat S. and Chaijaruwanich J..
Lanna Dharma Printed Character Recognition using k-Nearest Neighbor and Conditional Random Fields.
DOI: 10.5220/0004112801690174
In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2012), pages 169-174
ISBN: 978-989-8565-29-7
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
Table 1: Types of Lanna Dharma characters.
Figure 1: Writing example of Lanna Dharma word.
Figure 2: Example of confusion sets.
2 RELATED WORKS
Methods and techniques for OCR are various.
Selecting which method and technique depends on
the characteristics of characters and documents.
Some methods are appropriate with some typed of
characters while some are not appropriate with
others.
The k-NN algorithm is a very well known
approach. It is the simplest approach of all machine
learning algorithms. A lot of works in character
recognition use the k-NN classifier. In (Holambe et
al., 2010), they provided a comparative study of
Devanagari handwritten and printed character and
numerals recognition using Nearest-Neighbor
classifiers. This method used gradient and curvature
based feature extraction method and compared all
the nearest neighbor classification methods. In
(Alkhateeb et al., 2009), k-NN was used for off-line
handwritten Arabic word recognition. The word was
segmented and divided into overlapping blocks.
Absolute mean values were computed for each block
of segmented words which constituted a feature
vector. Finally, the resulting feature vectors were
used to classify the word.
CRFs were introduced by Lafferty et al. (2001).
A conditional random field is simply a conditional
distribution p(y|x) with an associated undirected
graphical model structure. Within this model
dependencies among the input variables x do not
need to be explicitly represented, affording the use
of rich, global features of the input. For example, in
natural language tasks, useful features include
neighboring words and word bigrams, prefixes and
suffixes, capitalization, membership in domain-
specific lexicons, and semantic information (Sutton
and Mccallum, 2007). Moreover, CRFs have been
applied to many domains, including computer
vision, bioinformatics and word recognition. In word
recognition, the first research, which CRFs were
applied for handwriting recognition, was proposed in
(Feng et al., 2006). CRFs were used to recognize the
entire word without character segmentation and they
formulated recognition problem as a problem of
labeling observation sequences on a large-
vocabulary of words. Similarly, the word
recognition using CRFs was presented in (Shetty et
al., 2007), the method was based on segmentation of
word image into characters and identification of
lexicon with the highest probability. Moreover,
CRFs were applied to the same problem in other
researches such as in (Zhou et al., 2009).
3 THE PROPOSED METHOD
In this section, the Lanna Dharma printed character
recognition system is presented. It is comprising of
four main processes including: preprocessing,
segmentation, classification with k-NN, and
reclassification with CRFs. The whole process is
illustrated in Figure 3.
3.1 Preprocessing
In our experiments, we have collected document
images from E-60 Lanna Printed E-book Project
(Chaijaruwanich, 2010) which is a collection of 60
ancient Lanna books, printed between 1907-1947,
including the legends, poems, laws, etc. The
document images were scanned at 300 dpi and
KDIR2012-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
170
Figure 3: Lanna Dharma printed character recognition
process.
stored as gray-level images. After document
scanning, preprocessing operations will be applied to
the document images to enhance the images to be
the suitable format. Preprocessing is divided into
three steps:
Binarization
Noise removal
Skew correction
In the binarization step, Otsu’s method (Otsu,
1979) is used for converting grey level image to a
binary. In the noise removal step, noise is removed
by connected-component labeling which detects the
dusty pixels in the document images. If the size of
pixels are smaller than the defined threshold, they
are defined as noise and would be removed. Finally,
in the skew correction step, we detect the skew angle
of the document image using Hough Transform
(Duda and Hart, 1972). Then, skew correction is
done by rotating the document image referring to the
skewed angle.
3.2 Segmentation
After finishing preprocessing, we apply the
segmentation process which is divided into four
steps:
Line segmentation
Sub-word segmentation
Level segmentation
Character segmentation
First, we employ horizontal projection profile to
segment a page into separated lines. Second, vertical
projection profile is employed to segment the lines
by empty spaces into sub-words. Then, we use
horizontal projection profile again to find upper and
lower positions of a middle level for the level
segmentation. In this paper, we consider three levels:
upper level (i.e. levels 2, 3 in Figure 1), middle level
(i.e. level 1 in Figure 1), and lower level (i.e. levels
4, 5 in Figure 1). Finally, sub-words are segmented
into sequence of characters by using connected
component labeling and center of mass techniques.
Figure 4 shows an example of segmentation. As a
result, the output from this process is a sequence of
character images as shown in Figure 5 which is
ordered by writing system of Lanna Dharma word.
Figure 4: Example of segmentation, ① refers to upper
level, ② refers to middle level, and ③refers to lower
level.
Figure 5: Sequence of character images ordered by writing
system of Lanna Dharma word.
3.3 Classification with k-Nearest Neighbor
Generally, a classifier in built upon a sample of
character images called the training set. The
classifier decision is based on feature vector
extracted from the character image.
3.3.1 Feature Extraction for k-NN
In this paper, height, width, aspect ratio, area and
level from original size of character images are used
as a feature vector. In addition, the character images
are normalized into 36*36 pixels and they are
divided into 9*9 cells, the ratios of number of the
black pixels and area of each cell are also used as
features. So in total we have 86 features for each
character image.
3.3.2 Classification with k-NN
For k-NN classification, training patterns are data
LannaDharmaPrintedCharacterRecognitionusingk-NearestNeighborandConditionalRandomFields
171
points in d-dimensional space, where d is the
number of features. An unlabelled test pattern is
another point within the same space and classified
by considering the most frequently occurring class
among its k-most similar training patterns (Duda and
Hart, 1973). For our experiments, we defined k is 3
and the similarity measure for k-NN classification is
the Euclidian distance metric, defined between
feature vectors x and y as:
,
(1)
where f represents the number of features, smaller
distance values represent greater similarity.
3.4 Reclassification with Conditional
Random Fields
We formulate a recognition problem as a problem of
labeling observation sequences as in (Feng et al.,
2006), and (Shetty et al., 2007). Generally, x is a
random variable over data sequences to be labeled,
and y is a random variable over corresponding label
sequences. All components y
i
of y are assumed to
range over a finite label alphabet У. In this paper, x
range over Lanna Dharma character sequences and y
range over labels of those sequences, with У is the
set of possible Lanna Dharma character symbols as
in Table 1.
3.4.1 Feature Selection for CRFs
We use the following features: type of character as
defined in Table 1, level of character as defined in
section 3.2, and word boundary which are B
(Beginning character of word) and I (Internal
character of word). The example of Lanna Dharma
character sequence, their features, and their true
label are shown in Figure 6.
3.4.2 Classification with CRFs
The random variables x and y are jointly distributed,
but in a discriminative framework as CRFs, we
construct a conditional model p(Y=y|X=x) from
paired observation and label sequences. The
probability of a label sequence y = y
1
, y
2
,...,y
t
given
an observation sequence x = x
1
,x
2
,...,x
t
can be written
as
|
1
exp
,,
(2)
Figure 6: Example of Lanna Dharma character sequence,
their features and their true label. The first column is the
output from k-NN classification, the second column is the
type of character, the third column is the level of
character, the fourth column is the word boundary, and the
fifth column is the true Lanna Dharma character label.
where
,,
(3)
θ is the model parameter, Z(x) is a normalized factor
which sums all cases of label y′ and F(y,x,t) is the
sum of feature frequency at position t.
We define the feature function of the entire
Lanna Dharma character sequences and their
features x and the corresponding Lanna Dharma
character label y at position t as
,,
,
,
,
,
,,
,
,
(4)
where
,
,
,
is the transition feature
function of Lanna Dharma character sequences and
their features x and label y at position t-1 and t equal
to l′ and l, orderly in the whole states of Lanna
Dharma character labels.
,
is the state
feature function of the Lanna Dharma character label
at position t of the Lanna Dharma character
KDIR2012-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
172
sequence y and their features x.
,
, and
are
feature weights associated respectively with
,
and
estimated from the training data. Let θ = (λ, β).
We define
,
1i
f
atpositionto
isequalto
,
0otherwise
(5)
,
,
,
,
if
and
0otherwise
(6)
,
,
if
0otherwise
(7)
From equation (5),
,
represents whether x at
position t - k to t + k consisting of the Lanna
Dharma character subsequence n and the Lanna
Dharma character feature s. We consider a
reasonable local information pattern of k = ± 2
Lanna Dharma character surround the considering
position t. In equation (6), and (7), l and l′ are Lanna
Dharma character labels.
CRFs prediction is performed by finding the
most probable label sequence y* from the training
model given the observation or Lanna Dharma
character sequences and the Lanna Dharma
character features x from the testing data.
∗
|
exp
,,
(8)
The Viterbi algorithm (Lafferty et al., 2001) is
used as a dynamic programming algorithm to
generate the inferences in CRFs. This is to find the
most probable label y* for the Lanna Dharma
character sequence and their features x with highest
probability value.
The graphical model of our proposed CRFs are
shown in Figure 7. We consider a range of two
characters forward and backward from the indicated
position t. These relationships are also used for
training and testing data formats.
Figure 7: Graphical model of our CRFs model.
In our experiment, CRFs are implemented using
CRF++ 0.53 (Kudo, 2005) which is an open source
program providing a generalized CRFs learning and
testing platform. CRF++ have been successfully
used in a lot of reseach such as (Shoombuatong et
al., 2011) and (Subpaiboonkit et al., 2012).
4 EXPERIMENTS
Generally, the recognition system needs a large
database to train and test the system. Therefore, we
select Lanna poem documents which are parts of the
E-60 Lanna Printed E-book Project
(Chaijaruwanich, 2010) to identify a scope of Lanna
word, and collect a small database to train and test
the Lanna Dharma printed character recognition
system. The documents are selected with each page
consists of 900-1,200 characters and they vary in
levels of noise.
30 pages of document are randomly sampled for
training k-NN classifier. The training dataset
consists of 145 classes, we randomly select 30
sample images for each class. Totally, we have
4,350 sample images of training data for
classification with k-NN. Samely, training data for
CRFs classifier, 30 pages of document are randomly
sampled, including 1,983 sequences or 33,721
characters.
To test our method, we randomly select 10 pages
of documents, including 714 sequences or 11,425
characters. We compare our method with
recognition using k-NN without CRFs. Table 2
shows the accuracy result of recognition using k-NN
without CRFs, and using both k-NN and CRFs.
Table 2: The recognition accuracy rate.
Classifier Recognition Accuracy Rate (%)
k-NN 78.27
k-NN and CRFs 82.61
Recognition accuracy of our method in Table 2 is
higher than recognition using k-NN without CRFs.
The recognition errors from our method may be
caused by two sources: the broken and touching
characters, and too small training dataset for CRFs.
Broken and touching characters error refers to the
case that the broken and touching characters are
incorrectly segmented, then they are the causes of
recognition error. Small training dataset for CRFs
may cause a prediction error for unseen patterns.
LannaDharmaPrintedCharacterRecognitionusingk-NearestNeighborandConditionalRandomFields
173
5 CONCLUSIONS
In this paper, we propose Lanna Dharma printed
character recognition and focus on solving the
problem of confusion sets. After the preprocessing
and segmentation processes, we classify Lanna
Dharma printed character using k-Nearest Neighbor.
The output from classification using k-Nearest
Neighbor is reclassified again by CRFs. To use
CRFs, we define Lanna Dharma character sequences
which are output from classification using k-Nearest
Neighbor as the sequence labeling task, and we
construct a suitable template for the CRFs to train
and test the data. The experimental results show that
our method increases the recognition accuracy.
ACKNOWLEDGEMENTS
The research described in this paper was supported
by the Graduate School Chiang Mai University,
Chiang Mai, Thailand.
REFERENCES
Alkhateeb, J. H., Khelifi, F., Jiang, J. and Ipson, S. S.
(2009). A new approach for off-line handwritten
Arabic word recognition using KNN classifier. Proc.
ICSIPA, 191 -194.
Chaijaruwanich, J. (2010). E-60 Lanna Printed Book
Project. Retrieved from: http://www.lannalit.org
Duda, R. O. and Hart, P. (1972). Use of the hough
transformation to detect lines and curves in pictures.
Comm. ACM, 11–15.
Duda, R. O. and Hart, P. (1973). Pattern classification and
scene analysis. Wiley, New York.
Feng, S., Manmatha, R. and Mccallum, A. (2006).
Exploring the use of conditional random field models
and hmms for historical handwritten document
recognition. Proc. 2nd DIAL, 30-37.
Holambe, A. N., Holambe, S. N. and Thool, R. C. (2010).
Comparative study of Devanagari handwritten and
printed character & numerals recognition using
Nearest-Neighbor classifiers. Proc. 3rd ICCSIT, 426-
430.
Kudo, T. (2005). CRF++: Yet another CRF toolkit.
Retrieved from: http://code.google.com/p/crfpp/
Lafferty, J., Mccallum, A. and Pereira, F. (2001).
Conditional random fields: Probabilistic models for
segmenting and labeling sequential data. Proc. 18th
ICML, 282-289.
Otsu, N. (1979). A threshold selection method from gray-
level histograms. IEEE Trans. on System, Man, and
Cybernetics, 9, 62–66.
Shetty, S., Srinivasan, H. and Srihari, S. (2007).
Handwritten word recognition using conditional
random fields. Proc. 9th ICDAR, 2, 1098 - 1102.
Shoombuatong, W., Traisathit, P., Prasitwattanaseree, S.,
Tayapiwatana, C., Cutler, R. and Chaijaruwanich, J.
(2011). Prediction of the disulphide bonding state of
cysteines in proteins using conditional random fields.
International Journal of Data Mining and
Bioinformatics, 5, 449-464.
Subpaiboonkit, S., Thammarongtham, C. and
Chaijaruwanich, J. (2012). RNA Family Classification
Using the Conditional Random Fields Model. Chiang
Mai Journal of Science, 39.
Sutton, C. and Mccallum, A. (2007). Introduction to
Statistical Relational Learning. MIT Press.
Zhou, X. D., Liu, C. L. and Nakagawa, M. (2009). Online
Handwritten Japanese Character String Recognition
Using Conditional Random Fields. Proc. 10th ICDAR,
521-525.
KDIR2012-InternationalConferenceonKnowledgeDiscoveryandInformationRetrieval
174
