Optimal Clustering Technique for Handwritten Nandinagari Character Recognition

International Journal of Computer Applications Technology and Research
Volume 6–Issue 5, 213-223, 2017, ISSN:-2319–8656
www.ijcat.com 213
Optimal Clustering Technique for Handwritten
Nandinagari Character Recognition
Prathima Guruprasad
Research Scholar, UOM,
Dept. of CSE, NMIT,
Bangalore, India
Prof. Dr. Jharna Majumdar
Sc. G DRDO (Retd.), Dean,
R&D, Prof. and Head,
Dept. of CSE and Center for Robotics Research,
NMIT, Bangalore
Abstract: In this paper, an optimalclustering technique for handwritten Nandinagari character recognition is proposed. We compare two
different corner detector mechanisms and compare and contrast various clustering approachesfor handwritten Nandinagari characters.
In this model, the key interest points on the images which are invariant to Scale, rotation, translation, illumination and occlusion are
identified by choosing robust Scale Invariant Feature Transform method(SIFT) and Speeded Up Robust Feature (SURF) transform
techniques. We then generate a dissimilarity matrix, which is in turn fed as an input for a set of clustering techniques like K Means,
PAM (Partition Around Medoids) and Hierarchical Agglomerative clustering. Various cluster validity measures are used to assess the
quality of clustering techniques with an intent to find a technique suitable for these rare characters. On a varied data set of over 1040
Handwritten Nandinagari characters, a careful analysis indicate this combinatorial approach used in a collaborative manner will aid in
achieving good recognition accuracy. We found that Hierarchical clustering technique is most suitable for SIFT and SURF features as
compared to K Means and PAM techniques.
Keywords: Invariant Features, Scale Invariant Feature Transform, Speeded Up Robust Feature technique, Nandinagari Handwritten
Character Recognition, Dissimilarity Matrix, Cluster measures, K Means, PAM, Hierarchical Agglomerative Clustering
1. INTRODUCTION
The awareness of very old scripts is valuable to historians,
archaeologists and researchers of almost all branches of
knowledge for enabling them to understand the treasure
contained in ancient inscriptions and manuscripts [1].
Nandinagari is a Brahmi-based script that was existing in India
between the 8th and 19th centuries. This is used as writing style
in Sanskrit especially in southern part of India. Nandinagari
script is older version of present day Devanagari script. But
there are some similarities between Nandinagari and
Devanagari in terms of their character set, glyphic
representation and structure. However, Nandinagari differs
from Devanagari in the shapes of character glyphs, absence of
headline. There are several styles of Nandinagari, which are to
be treated as variant forms of the script. Sri Acharya Madhwa
of the 13th century, a spiritual Leader who founded the Dvaita
school of Vedanta has hundreds of manuscripts written in
Nandinagari on the Palm leaves.
Nandinagari script is available only in manuscript form hence
it lacks the necessary sophistication and consistency. There are
innumerable manuscripts covering vast areas of knowledge,
such as Vedas, philosophy, religion, science and arts preserved
in the manuscript libraries in digital form. Today though
Nandinagari script is no longer in trend, the scholars of Sanskrit
literature cannot be ignorant of this script. Nandinagari
character set has 15 vowels and 37 consonants, 52 characters
as shown in Table 1 and Table 2. We face many challenges to
interpret handwritten Nandinagari characters such as
handwriting variations by same or different people with wide
variability of writing styles. Further, these documents are not
available in Printed Format and only handwritten scripts are
available. Absence of any other published research methods
using these rare characters makes if more challenging.
Nandinagari Optical Character Recognition (OCR) is not
available to date. Therefore, we need to extract invariant
features of these handwritten characters to get good recognition
accuracy.
Table 1. Nandinagari Vowels and Modifiers
Vowels Modifiers Vowels Modifiers

www.ijcat.com 214
In this paper we extract features using Scale Invariant Feature
Transform (SIFT) [2] and Speeded Up Robust Feature (SURF)
transform techniques [7]. The SIFT and SURF features are
local and based on the appearance of the object and are
invariant to different sizes and orientations. They are also
robust to changes in illumination, noise and highly distinctive
with low probability of mismatch. From these features, a
dissimilarity matrix is computed. Then this is given as an input
to different clustering techniques to group similar characters.
The set of clustering mechanisms identified for these characters
are K Means, PAM and Hierarchical agglomerative clustering
technique. The performance of these techniques are compared
and best method for SIFT and SURF features is identified.
Table 2. Nandinagari Consonants
2. RELATED WORK
The scale and variety of applications using SIFT [3][4][5][6]
and SURF[9] is discussed in many papers on pattern
recognition. The robustness of SIFT and its comparison with
SURF algorithm is also discussed in some papers [8][9]. The
recognition of multiple type of words including Devanagari
using Visual bag of words is discussed using SIFT algorithm
[11]. An attempt to classify human faces using SIFT and
hierarchical clustering approach is also introduced
[12].Clustering algorithms like K Means and K Medoids and
their performance is also discussed [14].Different cluster
measures to evaluate the performance of cluster methods are
also discussed [15].
3. METHODOLOGY
Handwritten Nandinagari character database is created
manually as standard dataset is not available. For a set of 52
vowels and consonants in Nandinagari, with an average of 5
different variations over the format of representation(jpg or
png), size(256X256, 384X384, 512X512, 640X640), degree of
rotation(0, 45, 90, 135 and 180 degree) and translation( positive
or negative offset of 15 pixels), a database of 1040 characters
isprepared. The proposed architecture shown in Figure. 1
consists of following steps:
Algorithm Steps
1. In the first step, all the characters in the database are
scanned.
2. In the pre-processing step, we convert these images into
their grayscale form.
3. Interest points from the input image are extracted using
the Scale invariant feature transform (SIFT) technique. From
each point, 128 feature descriptors are extracted which are
invariant to scale, rotation and illumination. Similarly, features
are also extracted using Speeded Up Robust Feature (SURF)
transform technique. 64 feature descriptors are generated from
each candidate points.
4. For each image in the database, the number of match
points are found with every other image and vice versa.
5. The maximum number of match points is computed by
considering number of match points and N X N match matrix
is generated.
6. The dissimilarity ratio is now computed using the
following formula
Eij = Eji = {100 * (1 - nMax / nMin) }, where nMax
= maximum number of match points in either directions and
nMin = minimum number of key points in either direction
7. The SIFT and SURF features dissimilarity matrix is fed
as input for different clustering techniques to group similar
handwritten Nandinagari characters together.
8. The best-suited clustering technique for SIFT and SURF
features is identified by analyzing the performance using
cluster measures.

www.ijcat.com 215
Figure 1. Proposed Model Architecture
3.1 Clustering
Clustering is the process of grouping a set of similar objects
into same clusters and dissimilar objects in other clusters. Three
prominent approaches are taken for analysis and comparison
here. They are K Means, PAM and Agglomerative Hierarchical
Clustering.
3.1.1 K Means Clustering
K-means clustering algorithm uses an iterative refinement
approach. Here we partition of the characters into k clusters,
such that the characters in a cluster are more similar to each
other than to characters in different clusters [14].This is based
on the Euclidean distance measure, we calculate the new mean
which is the centroid of the clusters and assign nearest points
and this process is continued until the cluster centres remains
unchanged.
3.1.2 PAM Clustering
This method chooses a character from all characters in the
dataset as medoids of a cluster i.e., a cluster centre, in contrast
to the K-Means method, which selects a random value as the
centre of the cluster. The objective is to minimize the average
dissimilarity of characters to their closest selected character.
This method starts from an initial set of medoids and iteratively
replaces one of the medoids by one of the non-medoids if it
improves the total distance of the resulting clustering [15]. The
PAM method is more robust to noise and outliers, compared to
the K-means method.
3.1.3 Agglomerative Clustering
Agglomerative Hierarchical Clustering is a bottom up approach
where each observation starts in its own cluster, and pairs of
clusters are merged as one moves up in the hierarchy [13].The
result of the hierarchical methods is a dendrogram, representing
the nested grouping of objects. There are different methods for
agglomeration such as single, complete, average methods. In
this paper, we have used the average linkage method as an
algorithm for this approach. This is better than the K Means
and PAM approaches since it automatically detects the number
of clusters.
3.2 Cluster Validation Measures
Choosing appropriate clustering method for a given dataset is a
very challenging task. So different clustering measures are
considered to validate the clustering results. It is helpful to
choose best clustering technique for a specific application.
Here we validate the results using two categories of measures
such as internal and stability validation measures. Internal
measure use the fundamental information in the data to evaluate
the quality of the clustering. Stability measure evaluate the
consistency of the clustering results [16].
3.2.1 Internal Measures
For internal measures, three measures are considered such as
Connectivity, Silhouette width and Dunn Index. Connectivity
is used to measure the connected component, which relates to
what extent items are placed in the same cluster as their nearest
neighbours in the data space. In the second measurement
approach, the silhouette value measures the degree of
confidence in the clustering assignment of a particular item.
This value ranging between -1 to 1 need to be maximized.
However, in the third measurement approach, the Dunn index
indicates the ratio of the smallest distance between items not in
the same cluster to the largest intra-cluster distance. The Dunn
index has a value between zero and one, and need to be
maximized.
3.2.2 Stability Measures
The stability measure compare the results from clustering based
on the original data set to clustering based on deleting one
column at a time. These measures work well if the data are
highly correlated. The stability measures considered here are
the average proportion of non-overlap (APN), the average
distance (AD), the average distance between means (ADM),
and the figure of merit (FOM). The APN measures the average
proportion of observations not placed in the same cluster by
clustering based on the original data and clustering based on
the data with a single column removed. The AD measure
computes the average distance between observations placed in
the same cluster by clustering based on the original data and
clustering based on the data with a single column removed.
The ADM measure computes the average distance between
cluster centres for observations placed in the same cluster by
clustering based on the original data and clustering based on
the data with a single column removed. The FOM measures the
average intra-cluster variance of the observations in the
removed column, where the clustering is based on the
remaining samples. This estimates the mean error using
predictions based on the cluster averages. The APN has the
value between 0 and 1 and with values close to zero
corresponds to highly consistent clustering results. Remaining
measures have values between zero and ∞and smaller values
are favoured for better clustering performance.
4. EXPERIMENTAL RESULTS
The results are obtained for various stages of character
recognition. The samples of images of different size 256 X 256,
384 X 384, 512 X 512,640 X 640, different orientation angles

www.ijcat.com 216
0o
, 45o
, 90o
, 135o
, 180o
are taken. This forms a 1040 character
in the database. All the 1040 characters are considered for
computation and for the sake of depicting the results of cluster
formation, we take a subset of 16 distinct characters from this
set.
4.1 Cluster using SIFT Features
For K-means clustering approach, the parameter to be set prior
to clustering is the number of clusters. The optimal number of
clusters i.e. 14 for SIFT features is derived using Elbow method
is as shown in Fig 2. The clusters obtained using this technique
is indicated in Appendix 1.
As seen in this figure, the instances are misclassified and
hence would yield a low accuracy rate.
Figure 2. Optimal Number of Clusters (14) using K-means
for SIFT features
For PAM the optimal number of clusters need to be mentioned
but the partition done around the medoids and this is better
compared to K Means approach. The cluster results using SIFT
features are as shown in Appendix 2.
The optimal number of clusters for PAM is same as K Means
method as shown in Fig 3.The dendogram after hierarchical
clustering using SIFT features for sample characters partitioned
automatically into 16 clusters is as shown in Appendix 3.
Figure 3. Optimal Number of Clusters (14) using PAM for
SIFT features
4.1.1 Cluster Validation Measures for SIFT Features
The internal validation measures are the Connectivity,
Silhouette width, Dunn Index derived for three different
clustering techniques, K Means, PAM, and agglomerative
hierarchical clustering techniques using SIFT features. The
clustering validation results are analysed and the optimal score
for these three measures as shown in Table 3. For internal
measures using SIFT features, hierarchical clustering with two
clusters performs better for connectivity measures. For Dunn
Index and Silhouette Width, hierarchical clustering with
fourteen clusters performs better. For good clustering, the
connectivity is minimized, while both the Dunn index and the
silhouette width is maximized. So from table 3 it appears that
hierarchical clustering performs better compared to the other
clustering techniques for each internal validation measure.
Table 3. Internal and Stability cluster validation measures
for SIFT Features
Internal Measures
Measures Value Cluster
Method
No. of
Clusters
Connectivity 8.5079 Hierarchical 2
Dunn Index 0.9191 Hierarchical 14
Silhouette
Width
0.7140 Hierarchical 14
Stability Measures
Method
No. of
Clusters
APN 0.0195 Hierarchical 14
AD 61.1687 Hierarchical 14
ADM 7.6345 Hierarchical 14
FOM 8.8808 Hierarchical 14
The graphical representation of the connectivity, Dunn
index, and Silhouette Width measures are as shown in
Fig.4 to 6.

www.ijcat.com 217
Figure 4. Graphical representation of the connectivity internal
measure using SIFT Features
Figure 5. Graphical representation of the Dunn index internal
Figure 6. Graphical representation of the Silhouette Width
internal measure using SIFT Features
The stability measures for K Means, PAM and agglomerative
hierarchical clustering techniques using SIFT features are
computed. The optimal scores of the measures such as APN,
AD, ADM, and FOM are as shown in Table 3. For better
clustering results the measures are minimized. From the table
3, for these measures, hierarchical clustering with fourteen
clusters gives the best score. The graphical representation of the
stability measures for SIFT features such as APN, AD, ADM
and FOM as shown in Fig.7 to Fig. 10.
Figure 7. Graphical representation of the ADM stability

www.ijcat.com 218
Figure 8. Graphical representation of the AD stability measure
using SIFT Features
Figure 9. Graphical representation of the APN stability
Figure 10. Graphical representation of the FOM stability
4.2 Clustering using SURF Features
For K-means clustering technique, the number of clusters are
decided by Elbow method for SURF features. The k value is
found as 16 as shown in Fig 11. The SURF features grouped
together using this approach is as shown in Appendix 4. The
misclassification rate is more in K Means method. For PAM
the optimal number of clusters is generated using Elbow
method and as shown in Fig 12.
PAM is better compared to K Means approach because
partition is done around the medoids which leads to low error
rate. The cluster results using SURF features is as shown in
Appendix 5.
Figure 11. Optimal Number of Clusters (16) using K-Means
for SURF features
Fig. 12. Optimal Number of Clusters (16) using PAM for
SURF features
The dendogram of hierarchical clustering using SURF features
for sample characters partitioned automatically into 16 clusters
is as shown in Appendix 6.
4.2.1 Cluster Validation Measures for SURF Features
The internal and stability cluster validation measures for SURF
features is used to evaluate the results of K Means, PAM and
agglomerative Hierarchical clustering methods.
The analysis is as shown in table 4 for different cluster
measures. Internal clustering validation, which use the internal
information of the clustering process to evaluate the efficiency
of a clustering method. It can be seen that for SURF features
among three clustering methods, hierarchical clustering with 2

www.ijcat.com 219
clusters performs better for Connectivity and with 16 clusters
for Dunn Index and Silhouette Width.
Clustering stability validation evaluates the consistency of a
clustering result by comparing it with the clusters obtained after
each column is removed, one at a time. It is analysed that for
SURF features, Hierarchical clustering with 16 clusters proved
to be better for APN, AD, ADM, FOM stability measures.
Table 4. Internal and Stability cluster validation measures
for SIFT Features
Internal Measures
Method
No. of
Clusters
Connectivity 8.5079 Hierarchical 2
Dunn Index 3.0797 Hierarchical 16
Silhouette
Width
0.8153 Hierarchical 16
Stability Measures
Method
No. of
Clusters
APN 0.0000 Hierarchical 16
AD 34.2118 Hierarchical 16
ADM 0.0000 Hierarchical 16
FOM 3.4201 Hierarchical 16
The corresponding graphical representation of these measures
as shown in Fig.13 to 19.
Figure 13. Connectivity internal measure for SURF Features
Figure 14. Dunn Index internal measure for SURF Features
Figure 15. Silhouette Width internal measure for SURF
Features
Fig. 16: ADM stability measure for SURF Features

www.ijcat.com 220
Fig. 17: AD stability measure for SURF Features
Fig. 18: APN stability measure for SURF Features
Fig. 19: FOM stability measure for SURF Features
5. CONCLUSION
The proposed Nandinagari character retrieval system based on
data visualization method and is highly scalable. The SIFT and
SURF methods detect the interest points and derives feature
descriptors. This approach requires no or minimal pre-
processing of images and still can identify images in varying
states of occlusion. Our main aim is to provide efficient and
robust descriptors which are then used to compute dissimilarity
matrix. SIFT descriptors are more robust compared to SURF
descriptors. But computation time for SURF is less compared
to SIFT method. Then dissimilarity matrix of these descriptors
are subjected to different clustering approaches to group similar
handwritten Nandinagari characters together. Prerequisite for
K-Means and PAM is to specify the number of clusters.
Performance of PAM is better compared to K Means.
Agglomerative clustering method is more suitable for both
SIFT and SURF descriptors. Further we can explore the
performance of these descriptors using wide variety of
clustering techniques.
6. REFERENCES
[1] P. Visalakshi, "Nandinagari Script", DLA Publication,
First Edition, 2003.
[2] D. G. Lowe, “Distinctive image features from scale-
invariant key points”, IJCV, vol. 60, no. 2, pp. 91–110,
2004.
[3] Mortensen, E.N., Deng, H., Shapiro, L., "A SIFT
descriptor with global context", In Computer Vision and
Pattern Recognition (CVPR 2005), 20-25,IEEE, Vol. 1,
184-190, June 2005..
[4] Ives Rey-Otero, Jean-Michel Morel and Mauricio
Delbarcio, “An analysis of Scale-space sampling in
SIFT”, ICIP, 2014.
[5] Yishu Shi, Feng Xu, Feng-Xiang Ge Yishu Shi, Feng Xu,
Feng-Xiang Ge “SIFT-type Descriptors for Sparse-
Representation Based Classification”, 10th International
Conference on Natural Computation IEEE 2014.
[6] D. G. Lowe, “Object recognition from local scale-
invariant features,” ICCV, 1999.
[7] Baofeng Zhang, Yingkui Jiao Zhijun Ma, Yongchen Li,
JunchaoZhu,"An Efficient Image Matching Method Using
Speed Up Robust Features", International Conference on
Mechatronics and Automation IEEE 2014.
[8] Dong Hui, Han Dian Yuan, "Research of Image Matching
Algorithm Based on SURF Features", International
Conference on Computer Science and Information
Processing (CSIP) , IEEE 2012.
[9] Goh K M, Abu-Bakar S A R, Mokji M M, et al., "
Implementation and Application of Orientation
Correction on SIFT and SURF Matching",IJASCA, vol.5,
no.3, 2013.
[10] Martin A. Fischler and Robert C. Bolles, “Random sample
consensus: a paradigm for model fitting with applications
to image analysis and automated cartography,” in Journal
on Communications of the ACM, ACM Press, NY Vol 24,
No. 6, pages 381 to 395, 1981.
[11] Ravi Shekhar, C.V. Jawahar, “Word Image Retrieval
using Bag of Visual Words” IEEE, DOI
10.1109/DAS.2012.96, 2012.
[12] Panagiotis Antonopoulos, Nikos Nikolaidis and Ioannis
Pitas, “Hierarchical Face Clustering Using Sift Image

www.ijcat.com 221
Features”, Computational Intelligence in Image and
Signal Processing, IEEE Symposium on April 2007.
[13] Akanksha Gaur, Sunita Yadav, "Handwritten Hindi
Character Recognition using KMeans Clustering and
SVM", 2015 4th International Symposium on Emerging
Trends and Technologies in Libraries and Information
Services, IEEE 2015.
[14] Hae-Sang Park, Jong-Seok Lee, Chi-Hyuck Jun , " A K-
means-like Algorithm for K-medoids Clustering and Its
Performance ", Proceedings of the 36th CIE Conference
on Computers & Industrial Engineering, Page No 1222-
1231, 2006.
[15] Oliver Kirkland, Beatriz De La Iglesia, " Experimental
Evaluation of Cluster Quality Measures", IEEE, 2013.
[16] Guy Brock, VasylPihur, SusmitaDatta, SomnathDatta,"
clValid: An R Package for Cluster Validation " Journal of
Statistical Software, Vol 25, Issue 4, March 2008.
Appendix 1: Clustering SIFT features with K Means method for sample characters (14 clusters)

www.ijcat.com 222
Appendix 2: Clustering SIFT features with PAM (Partition around Medoids) for sample characters (14 clusters)
Appendix 3: Clustering SIFT features with Agglomerative Hierarchical clusters for sample characters (16 clusters)

www.ijcat.com 223
Appendix 4: Clustering SURF features with K Means method for sample characters (16 clusters)

www.ijcat.com 224
Appendix 5: Clustering SURF features with PAM (Partition around Medoids) for sample characters (16 clusters)
Appendix 6:. Clustering SURF features with Agglomerative Hierarchical clusters for sample characters (16 clusters)

Optimal Clustering Technique for Handwritten Nandinagari Character Recognition

More Related Content

What's hot (19)

Similar to Optimal Clustering Technique for Handwritten Nandinagari Character Recognition (20)

More from Editor IJCATR (20)

Recently uploaded (20)

Optimal Clustering Technique for Handwritten Nandinagari Character Recognition