Proposal for Development of Online Handwritten Character Recognition Engines for Indian languages

1
Proposal for Development of Online Handwritten
Character Recognition Engines for Indian
languages

• V. Srinivasa Chakravarthy1 C. Chandrasekhar2
• 1Biotechnology Department, 2Department of
  Computer Science,Indian Institute of Technology,
  Madras,Chennai 600036.

2
Objectives

• To develop Online Handwritten Character
  Recognition Engines for
• Hindi
• Tamil
• Telugu
• Malayalam
• Kannada

2
3
Approach

• Stroke Recognition Level
• Support vector machines
• Shape Feature Analysis
• Character Recognition Level
• Rule-based system
• Finite State Automaton

3
4
Issues

• Composite characters
• Variability for same character
• Minor variations in similar characters
• Large number of character and stroke classes

Same character
Different characters that look similar
4
5
Process Flow
Pre- processing
Feature Extraction
Stroke Identification
Character Recognition
Stroke Stream
Char Code
6
Preprocessing
7
Feature extraction
8
Stroke Classification
9
Character Identification
a b c d c e f d g h d i j h k
l j m n
10
Devanagari Script

• Training
• 90 users, 21780 examples
• Testing
• 10 users, 2420 examples
• 115 stroke classes.

• Stroke level classification 91.53.
• ? 20, C 100.

11
Telugu Script

• Training 82 users, 33726 samples
• Testing 10 users, 4091 samples
• 253 stroke classes
• Classification accuracy 83.08
• C 50, ? 30
• Pre-classification of strokes
• Based on stroke position within the character
• Based on the presence of a loop with
  x-intersection.

12
Telugu Script

• Strokes touching/just above the base line
• having loop with x-intersection

• Strokes touching/just above the base line
• not having loop with x-intersection.

• Strokes below the base line

• Strokes well above the base line top strokes.

• Performance with 4 categories

13
Feature Extraction
Feature String for the above stroke
is CLQLIPONMLKJIPA
14
Feature String Matching
X
ABC
XXX
XXXX
DEFG
H
XXXXX
DEFG
XXX
H
ABC
XX
15
Character Recognition
State Transition Diagram for processing
sequences.
16
Performance
17
Tamil Example
18
Malayalam Example