Vous êtes sur la page 1sur 50

Presented by:

Shashi Kiran.S
Lecturer, JNN College of Engineering,
Shimoga, Karnataka

Under the Guidance of :
Dr.R.Srinivasa Rao Kunte
Principal, JNN College of Engineering
Shimoga, Karnataka
Introduction
On-line Character Recognition System
Literature survey
Problem statement
Input system
Pre-processing
Feature extraction
Classifier
Structural features
Software development
Results
Conclusion
List of publications


Today's computers are being used in all walks of
our life

Great Emphasis is for interaction with Regional
Languages

Indian Languages are rich in alphabet

Example : HINDI has more than 600 characters


11 vowels


36 consonants

Example of Halant form of consonants
Example of Conjunct characters
It is tedious and cumbersome to type Hindi text using English
converted Hindi keyboard.

It is very convenient and faster to enter a Hindi document into
the computer by natural way of handwriting, rather than by
typing using the existing English converted Hindi keyboard,
where key combinations have to be used.

This calls for an efficient On-line Hindi character recognition
system:

- Hindi Characters are written on a Digitizer Tablet with its
Electronic Pen (like a conventional pen-pad system).

- The Recognition system captures the character sequence
on-line, extracts features and recognizes/classify it and
displays it in Hindi Computer font.



Substantial variation in Size & Shape
Speed of writing
Inclination of characters





Digitizer tablet can be used for writing the text
Character contour is captured as (x,y) co-ordinates
Character is pre-processed for noise & size
variation elimination etc.
Features are extracted from pre-processed
character
Classifiers are used to recognize and classify
characters using features of characters
Good progress in last two decades
- due to:
More powerful computers and accurate
electronics equipment scanners, cameras
and electronic tablets

Survey done on research work on Bangla, Arabic,
persian, English, Nepalese, Kannada, Tamil,
Telugu.. Languages

Tools used HMM, ANN, SVM, BPM ..
Most of the research work are focused on foreign
languages

Very less research works on Indian languages

Survey done on research work on Hindi character
recognition:

- no work is reported for the recognition
methodology of complete set of Hindi characters

- substantial scope for the development of efficient
system using simple set of character features and
efficient classifiers


To develop an efficient On-line Hindi character
recognition system which

Should capture the character sequence written
on a Digitizer pad
Should be able to recognize vowels and
consonants and to further demonstrate that it
can even recognize the consonant vowel (CV)
combinations
Should display the recognized handwritten
character in Hindi font
Should be fairly writer independent

Digitizer tablet is used to write characters

Character contour is captured as a set of
digitized data points as (x, y) coordinates of
the pen contacts on the pad

Raw (x, y) coordinates from Digitizer pad

Filtering is done to eliminate noise due to
Hand fluctuation while writing
Errors in the sensing mechanism
Before Filtering After Filtering
Character written can vary in size
Horizontal and vertical size normalizations
No. of points captured for a character varies with
respect to its size, shape & speed of writing
Interpolation
After Re-sampling Before Re-sampling

To capture the essential characteristics of the
character

Features extracted represent a character for
recognition purpose

Features are separately extracted for both x and y
coordinates of character

They are together used as the features of the character
for recognition


Feature Extraction
Fx = Dwt(X); Fy = Dwt(Y)
F = Fx+Fy
Fx=Dwt(X)
Fy=Dwt(y)
Fx
Fy
F=Fx + Fy

The task of a classifier is to use the feature vector
provided by the feature extractor to assign the
object to a category

Identification of characters

Method used- Artificial Neural Networks
Set of characters considered for recognition
Phase-1 : Vowels classifier
Classifier designed to recognize 36 Hindi consonants
Sometimes misclassification b/w the following
pair of Similar characters by the Neural network

If Classifier output happens to be one of the
characters as shown in previous Fig., then the
third stage of Structural classification is
invoked.

Set of rules have been defined for each pair
structure for discrimination between them.
Similarity b/w and
Structural Features contd..
Once the character is recognized by NN, the
networks number is mapped to the
corresponding Hindi character font from
Character Map of the font

Font BRH Devnagari from Baraha font is
used to display the Hindi Font

Visual Basic is used to develop Editor window

Neural network toolbox of MATLAB is used to
create, train the network and recognize the
characters

Wavelet toolbox is used to extract features
using wavelets

1. The Visual Basic program part to display the Editor
window is executed which displays the Editor
window
2. The MATLAB GUI program is executed to display the
GUI window. Now, the display will have both Editor
window and GUI window.
3. Click the CLEAR push button to clear the
handwritten character in the WRITING AREA and
displayed font in RECOGNISED CHARACTER text
box, if any.
4. Write a character in the WRITING AREA of Editor
window.
5. Click the RECOGNISE push button to execute the
Character Recognition program
Character_Recognition.m which identifies and
recognize the handwritten character.
6. Click the DISPLAY push button to display the
recognized character in the Hindi font.


Experiments were carried out to fix
The number of nodes in the Hidden layer
No. of training samples to be used for
training the network
Performance of the system is tested by
varying nodes of hidden layer

52 number of nodes in Input layer, 10 in
hidden layer and 11 in output layer for
recognizing 11 Vowels were considered

30 samples were used for each class of
character for training
Training didnt
converge at all
0.00
20.00
40.00
60.00
80.00
100.00
120.00
20 30 40 50
Hidden layer nodes
R
e
c
o
g
n
i
t
i
o
n

r
a
t
e
0
20
40
60
80
100
120
5 10 15 20 25 30
Number of samples
R
e
c
o
g
n
i
t
i
o
n

r
a
t
e
The system is both size and
position invariant
Database used in the experiment consists of
characters obtained from various people

Samples are collected from 6 persons

30 samples are taken for each class

Methodology developed to recognize the
complete set of Hindi Characters:

- Two levels of neural networks are used in
cascade: Main & Sub neural N/Ws

System is both Size & Position invariant




System is able to recognize even if the
character is written in slightly varying angles

The system is able to distinguish and identify
the characters which are similar to one
another in structure

System can be made adaptive to a specific
user by training the system with his/her
writing
The system methodology can be extended to
recognize Consonant-Vowel (CV) characters
of remaining consonants to

Suitable methodology can be developed for
recognizing Conjunct Hindi Characters.
International conference
Shashi Kiran.S and R.Srinivasa Rao Kunte, On-line
handwritten Hindi Character recognition using Artificial
Neural Networks. Third International Conference on
Information Processing, August 2009, Bangalore,
Karnataka, India, pp 53-61.

Shashi Kiran.S and R.Srinivasa Rao Kunte, On-line
handwritten Hindi Character recognition System, APSIPA
Annual Summit and Conference, Dec14-Dec17, 2010 at
Biopolis, Singapore.
(selected but could not be registered in time)

Shashi Kiran.S and R.Srinivasa Rao Kunte, On-line
handwritten Hindi character recognition using
Artificial Neural Networks and structural features.
National Conference on Signal Processing
Communications and VLSI design (NCSCV), May
2009, Coimbatore, Tamilnadu, India, pp 80-81.

Shashi Kiran.S and R.Srinivasa Rao Kunte, A novel
cascade ensemble classifier system for on-line
handwritten Hindi character recognition. National
conference on Image processing, Pattern
recognition and computer vision (NCIPV), August
2009, Mandya, Karnataka, India. pp 18.

Vous aimerez peut-être aussi