DESIGN AND DEVELOPMENT OF AN ONLINE CHARACTER RECOGNITION SYSTEM FOR MALAYALAM

1
DESIGN AND DEVELOPMENT OF AN ONLINE CHARACTER
RECOGNITION SYSTEM FOR MALAYALAM

• Language Technology Centre
• C-DAC, Thiruvananthapuram

2
Proposer

• R. Ravindra Kumar
• Senior Director Head
• Language Technology Centre
• C-DAC, Thiruvananthapuram

Institution
Centre For Development of Advanced Computing
(C-DAC) Vellayambalam Thiruvananthapuram 695
033
3
Language - Malayalam

• Components that will be implemented
• Core Engine for Malayalam On-line Character
  Recognition

4
Techniques to be used - Proposed System

• System based on Multi-Layer Perceptron (MLP)
  Neural Network with back-propagation learning
  algorithm.
• Back-propagation learning algorithm
• For fast processing.
• Good performance in pattern recognition
• Windowing technique used to solve the
  limitations of MLP and to avoid the distortion
  caused by the normalization process.
• Makes use of both local and global features of
  characters.

5
MLP-Neural Network

• The multi-layer perceptron is an artificial
  neural network that learns nonlinear function
  mappings.
• Capable in supervised pattern matching.
• Learn by adapting its activation functions in
  order to match the input patterns to output
  patterns.
• In the proposed system
• Input nodes - feature vectors.
• Output nodes - corresponding character.

6
Windowing Technique

• To solve the limitation of MLP (fixed number of
  input neurons) and to avoid the distortion caused
  by the normalization process.
• To divide the sequence of features into
  components with overlapped frames.
• Each window contains equal number of features and
  acts as a MLP input neuron.
• Window size determined empirically.

7
System Overview
ONLINE CHARACTER RECOGNITION SYSTEM FOR MALAYALAM
8
Preprocessing

• Slant Correction
• To Normalize different slants of characters.
• De-hooking
• Process to remove the hooking stroke that may
  occur when writer downs or ups the pen on the
  tablet.
• Re-sampling
• Process to reduce unnecessary coordinate sequence
  data, received from the digital tablet.

9
Feature Extraction Character Recognition

• Extracts local and global features from the
  pre-processed data.
• Twelve features are extracted from each stroke.
• Six Global and Six Local features.
• Using the extracted features, the recognition
  engine based on MLP recognizes the character.

10
Performance of this Technique for other Languages
For free-hand Thai single characters, the
average recognition rate of 91.74 has been
achieved by using this method.1
11
Applications

• Word Processing
• Notes and jottings on Personal Digital Assistants
• Online Form Filling
• Teaching Aids
• Touch screen data input
• Diagnosis of Neuromuscular disorders
• Ticketing Machine

12
Schedule
13
References

• Online Handwritten Character Recognition Using
  Windowing Back-propagation Neural Networks (Sutat
  Sae-Tang Ithipan Methaste) International
  Conference on Modelling, Identification, and
  Control (MIC2002), Innsbruck, Austria, pp
  337-340, 18-21 Feb 2002.
•  
• Template Based Online Character recognition
  (Scott D Connell Anil K Jain) Pattern
  Recognition, Vol. 34 (1), pp. 1-14, Jan. 2001.
•  
• Recognition of unconstrained online Devanagari
  Characters (R M K Sinha, Scott D Connell Anil K
  Jain) Proceedings of the 15th International
  Conference on Pattern Recognition, Barcelona,
  Spain, pp. 368-371, Sept. 2000.
•  
• On line recognition of Hand written text Based on
  Hidden Markov Model. (Takashi Sudo) Thesis
  submitted to Japan Advanced Institute of Science
  and Technology

14

• Thank You