International Journal of Computer Science & Communication Vol. 1, No. 1, January-June 2010, pp.

91-95

Optical Character Recognition (OCR) for Printed Devnagari Script Using

Artificial Neural Network
Raghuraj Singh1, C. S. Yadav2, Prabhat Verma3, Vibhash Yadav4
1,3
Department of Computer Science & Engineering, Harcourt Butler Technological Institute, Kanpur-208002, India
2
Department of Computer Science and Engineering, Noida Institute of Engineering and Technology, Greater Noida-201306,
India
4
Pranveer Singh Institute of Technology, Kanpur-208016, India
Email: [email protected], [email protected], [email protected], [email protected]

ABSTRACT
There are about 300 million people in India who speak Hindi and write Devnagari script. Research in Optical Character
Recognition (OCR) is popular for its application potential in banks, post offices, defense organizations and library
automation etc. However most of the OCR systems are available for European texts. In this paper, we have proposed
a technique for OCR System for different five fonts and sizes of printed Devnagari script using Artificial Neural
Network. The recognition rate of the proposed OCR system with the image document of Devnagari Script has been
found to be quite high.
Keywords: OCR, Preprocessing, Segmentation, Feature Extraction, Classification, ANN, Skew Detection and Correction

1. INTRODUCTION preprocessing step it is expected to include noise removal,

Optical Character Recognition is a process by which we skew detection & correction. After finding out the feature
convert printed document or scanned page to ASCII of the segmented characters artificial neural network
character that a computer can recognize. The document (ANN) [1], [3] and [4] will be used for classification
image itself can be either machine printed or purpose. Efforts have been made to improve the
handwritten, or the combination of two. Computer performance of character recognition using artificial
system equipped with such an OCR system can improve neural network techniques. The proposed OCR system
the speed of input operation and decrease some possible shall be capable of accepting document images from a
human errors. Recognition of printed characters is itself file or from a scanner directly. Recognized characters can
a challenging problem since there is a variation of the also be displayed and edited.
same character due to change of fonts or introduction of
different types of noises. Difference in font and sizes 2. DESIGN OF OCR
makes recognition task difficult if preprocessing, feature Various approaches used for the design of OCR systems
extraction and recognition are not robust. There may be are discussed below:
noise pixels that are introduced due to scanning of the
image. Besides, same font and size may also have bold Matrix Matching: Matrix Matching converts each
face character as well as normal one. Thus, width of the character into a pattern within a matrix, and then
stroke is also a factor that affects recognition. Therefore, compares the pattern with an index of known characters.
a good character recognition approach must eliminate Its recognition is strongest on monotype and uniform
the noise after reading binary image data, smooth the single column pages.
image for better recognition, extract features efficiently, Fuzzy Logic: Fuzzy logic is a multi-valued logic that
train the system and classify patterns. Till now there is allows intermediate values to be defined between
no complete OCR for printed Devnagari Script which conventional evaluations like yes/no, true/false, black/
gives 100% success rate. white etc. An attempt is made to attribute a more human-
In this paper, we present a scheme to develop like way of logical thinking in the programming of
complete OCR system for different five fonts and sizes computers. Fuzzy logic is used when answers do not have
of Devnagari characters so that we can use this system a distinct true or false value and there is uncertainly
in Banking and Corporate sectors. We have implemented involved.
steps of the OCR system like preprocessing, Feature Extraction: This method defines each
segmentation, feature extraction and classification. In character by the presence or absence of key features,
92 International Journal of Computer Science & Communication (IJCSC)

including height, width, density, loops, lines, stems and inferior to the original, News papers: generally printed
other character traits. Feature extraction is a perfect on low quality paper etc.
approach for OCR of magazines, laser print and high
For such degraded documents, the system
quality images.
recognition accuracy comes down to 80-90%. But if we
Structural Analysis: Structural Analysis identifies want to use the OCR system for Banking and Corporate
characters by examining their sub features- shape of the sector, this accuracy rate is not up-to-mark.
image, sub-vertical and horizontal histograms. Its
Devnagari is most popular script to write Hindi as
character repair capability is great for low quality text
well as Sanskrit, Marathi, Sindhi, and Nepali language
and newsprints.
with minor modifications.
Neural Networks: This strategy simulates the way
the human neural system works. It samples the pixels in
each image and matches them to a known index of
character pixel patterns. The ability to recognize
characters through abstraction is great for faxed
documents and damaged text. Neural networks are ideal
for specific types of problems, such as processing stock
market data or finding trends in graphical patterns.

2.1. Structure of OCR Systems

Diagrammatic representation of the structure of an OCR
system is given in figure 1.

Fig 2: Stages in OCR Design

3. PROPOSED OCR SYSTEM

Following steps have been followed in the design of
proposed OCR system:
• Preprocessing;
• Segmentation;
• Feature Extraction;
• Classification.

3.1. Preprocessing
In the proposed OCR system, text digitization is done
by a flatbed scanner having resolution between 100 and
600 dpi. The digitized images are usually in gray tone,
and for a clear document, a simple histogram based
Fig. 1: Diagrammatic Structure of the OCR System. threshold approach is sufficient for converting them to
two tone images. The histogram of gray values of the
2.2. Stages in Design of OCR Systems pixels shows two prominent peaks, and a middle gray
value located between the peaks is a good choice for
Various stages of OCR system design are given in figure 2. threshold.

2.3. Reasons for Poor Performance of OCR Systems For salt and pepper noise we generally use median
filter. Median filter replaces the value of a pixel by the
Existing OCR systems generally show poor performance median of gray levels in the neighborhood of that pixel
for documents like old books: print and paper quality (the original value of the pixel is included in the
inferior due to aging, Copied Materials: documents like computation of the median), Median filters provide
photocopies or faxed documents, where print quality is excellent noise reduction capabilities, with considering
 Optical Character Recognition (OCR) for Printed Devnagari Script Using Artificial Neural Network 93

less blurring than linear smoothing filters of similar size

as shown in figure 3 & 4.

Fig 5: Vowels Modifiers of Devnagari

Fig 3: Image with Salt and Pepper Noise

Fig 6: Basic Characters of Devnagari

Fig 4: Image without Salt and Pepper Noise
As shown in figure 7, in Devnagari script, a text word
Derivative operator enhances edges and other may be partitioned into three zones. The upper zone
discontinuities (noise) and deemphasizes area with denotes the portion above the headline, the middle zone
slowly varying gray level values. covers the portion of basic and compound characters
below the headline, and the lower zone may contain
3.2. Segmentation where some vowel and consonant modifiers can reside.
For a long number of characters (basic as well as
Segmentation is one of the most important phases of OCR
compound) there exists a horizontal line at the upper
system. By applying good segmentation techniques we
part called “shirorekha” or headline in Hindi. The
can increase the performance of OCR. Segmentation
imaginary line separating the middle and lower zone
subdivides an image into its constituent regions or may be called the base line.
objects. Basically in segmentation, we try to extract basic
constituent of the script, which are certainly characters.
This is needed because our classifier recognizes these
characters only.
Segmentation phase is also crucial in contributing
to this error due to touching characters, which the
classifier cannot properly tackle. Even in good quality Fig 7: Partitioning of a Text Word into Zones
documents, some adjacent characters touch each other Line, Word and Character Segmentation: Once the
due to inappropriate scanning resolution. Numbers of text blocks are detected, the OCR system automatically
constituent characters touching each other in Devnagari finds individual text lines, segments the words, and then
and Bangla scripts are shown in table 1. separates the characters accurately.
Table 1 Segmentation of Line: Text lines are detected by
Constituent Characters Touching each other horizontal scanning. For segmentation of line, we scan
Script Touching An Image of Touching Characters of scanned document page horizontally from the top and
Characters Consists find the last row containing all white pixels, before a
black pixel is found. Then we find the first row containing
Two Three Four
entire white pixel just after the end of black pixels. We
Devnagari 11577 11183 394 nil repeated this process on entire page to find out all lines.
96.6% 3.4%
Segmentation of Words: After finding a particular
Bangla 16714 15277 953 484 line we separate individual words. This is done by
91.4% 5.7% 2.9%
vertical scanning.
To tackle the touching characters in Devnagari
Segmentation of Individual Characters: Once we
documents, at first, we attempt to identify the touching get the words we segment it to individual characters.
characters. Next, they are segmented into constituent Before segmenting words to individual characters, we
ones using a fuzzy decision making approach. locate the head line. This is done by finding the rows
Basic characters and vowels modifiers of the having maximum number of black pixels in a word. After
Devnagari are shown in figure 5 & 6. locating head line we remove it i.e. converts it in white
pixels. After removing head line our word is divided into
94 International Journal of Computer Science & Communication (IJCSC)

three horizontal parts known as upper zone, middle zone

and lower zone. Individual characters are separated from
each zone by applying vertical scanning.
Output of segmentation algorithm is shown in figure 8.

Fig 9(a): Output of Vertical Projection of Kha

Fig 8: Output of Segmentation

3.3. Classification
Fig 9(b): Output of Mean Feature Vector of Kha
Classification is performed based on the extracted
features. Here we are using ANN approach.
For initial classification of characters, we consider
three features as follows:
• Mean Distance;
• Histogram of projection based on spatial position
of pixel;
• Histogram of projection based on pixel value.
ANN Approach for Classification: Artificial Neural
Network approach has been used for classification and
recognition. It is a computational model widely used in
situation where the problem is complex and data is
subject to statistical variation. Training and recognition Fig 9(c): Histogram of KA Rowwise
phase of the ANN has been performed using
conventional back propagation algorithm with two 4. RESULTS AND DISCUSSIONS
hidden layers. The architecture of a neural network The experiments have illustrated that the artificial neural
determines how a neural network transfers its input into network concept can be applied successfully to solve the
output. This transfer can be viewed as a computation Devnagari Optical Character Recognition Problem. There
are many factors that affect the performance of OCR
3.4. Feature Extraction system for Devnagari Script. It is concluded that the input
matrix of size 48X57 gives better results than other
Feature extraction is one of the most important steps in choices. The recognition rate of OCR system with the
developing a classification system. This step describes image document of Devnagari Script is quite high as
the various features selected by us for classification of shown in the output.
the selected characters.
However, other kinds of preprocessing and neural
Classification based on the above three features has network models may be tested for a better recognition
been shown in figure 9(a), 9(b) & 9(c). rate in the future research in OCR System. Character
segmentation method which is incorporated in this paper
 Optical Character Recognition (OCR) for Printed Devnagari Script Using Artificial Neural Network 95

could be improved to handle large variety of touching Script Using Fuzzy Multi factorial Analysis”, IEEE
characters that occur often in images obtained from Transaction on System, Man and Cybernetics- Part C:
inferior-quality documents. The test set used in this Applications and Reviews, 32, November 2002. Page(s):
experiment is of 77 characters of five different types of 449-459.
fonts. This can be increased for better results. The [4] B. B. Chaudhary and U. Pal, “OCR Error Detection and
toughest phase in the experiment is getting a good set of Correction of an Inflectional Indian Language Script”,
characters for classification. Pattern Recognition 1996, IEEE Proceeding of 13 th
International Conference on 25-29 Aug., 3, 1996 page(s):
245-249.
5. FUTURE SCOPE OF WORK
[5] Nallasamy Mani and Bala Srinivasan, “Application of
Future enhancements that can be done on this paper
Artificial Network Model for Optical Character
include use of a dictionary of words to correct the output
Recognition”, System, Man and Cybernetics, 1997,
[8]. Implementing use of dictionary words may improve “Computational Cybernetics and Simulation”. 1997 IEEE
the performance of OCR system. One can also implement International Conference on 12-15 Oct. 1997 page(s): 2517-
the project for classifying hand-written text. 2520 3.
Segmentation of characters in hand written documents
[6] Veena Bansal and R.M.K. Sinha, “A Complete OCR for
is very complex as compared to printed documents. Multi
Printed Hindi Text in Devnagari Script”, Sixth
factorial Fuzzy System can be used for segmenting the
International Conference on Document Analysis and
characters in hand written documents.
Recognition, IEEE Publication, Seatle USA, 2001. Page(s):
800-804.
REFERENCES
[7] Veena Bansal and R.M.K. Sinha, “A Devnagari OCR and
[1] S. Mori et. al, “Historical Review of OCR Research and A Brief Overview of OCR for Indian Script”, PROC
Development”, Proceeding IEEE, 80, no 7, pp. 1029-1058,
Symposium on Transaction support System (STRANS 2001),
July 1992.
Feb. 15-17, 2001, Kanpur, India.
[2] A. A. Chaudhary, E.A.S. Ahmad, S. Hossain, C. M.
Rahman, “OCR of Bangla Character Using Neural [8] Bansal, V., Sinha, R.M.K., “Partitioning and Searching
Network: A better Approach”, 2nd International Conference Dictionary for Correction of Optically Read Devnagari
on Electrical Engineering (ICEE 2002), khuln, Bangladesh. Character Strings”, Document Analysis and Recognition,
[3] Utpal Garain and Bidyut B. Chaudhary, “Segmentation 1999. ICDAR’99, Proceedings of the Fifth International
of Touching Character in Printed Devnagari and Bangla Conference on 20-22 Sept. 1999 Page(s): 653-656.