OPTICAL CHARACTER RECOGNITION

1
OPTICAL CHARACTER RECOGNITION

• A proposal from IIIT Allahabad

2
Project Title

• Development of OCR technologies with
• application to various Indian and Tibetan
• script.

3
Objectives

• The two major objectives are
• To develop a toolset that will allow rapid
  prototyping of OCRs will allow experimentation
• To investigate the use of Gabor filters for
  feature extraction and test its applicability on
  various Indian and Tibetan scripts allows
  robust handling of degardation

4
Modules foreseen

• Annotation tool / Ground Truth Generator
• Development of general framework
• Visual component recognizer engine using Gabor
  filter
• Line component recognizer engine using HMM
• Tibetan OCR

5
Annotation tool / Ground truth generator

• To be used for generating the training sets for
  OCR in a short span of time and testing its
  accuracy for degraded images. The purpose of this
  tool is twofold.
• - It can be used to generate conveniently large
  amounts of ground truth data that would be
  suitable as training data for building an OCR
  with degradations
• - It can be used for testing an OCR and
  performing an analysis of its errors in the form
  of confusion matrixes useful for OCR fine tuning

6
Development of general framework

• Involves development of open architecture
• for optical character recognition
• Design consists of an OCR system that has
• It has component oriented design
• Highly configurable
• Any new algorithm can be added to system very
  easily

7
Design
8
Important issues

• Interface definition
• Control panel for configuring plugins
• Installation package

9
Visual component recognizer engine using Gabor
filter

• A complete visual component recognizer engine
  will be built that exploits the robustness of the
  Gabor filter feature set.
• The engine would accept visual components as
  input, extract estimates for line width, stroke
  width etc.
• Adjust its parameters automatically to produce
  parameters that give the best set of filters.
• These filters will then be used to obtain the
  feature vector corresponding to the given visual
  component.
• The engine will also provide various classifiers
  that would allow the final classification of the
  visual component.

10
Some observations

• Experiments with degraded text images show that
• the chief source of error is at the level of
• segmentation of characters.
• A similar situation exists for recognition of
  hand-
• written texts.
• Error rates are at acceptable levels for the
  other
• stages i.e. line segmentation, word segmentation,
• character recognition etc.
• Solution Develop a strategy that avoids
  character segmentation.

11
Line component recognizer engine using Gabor
filter

• To build an OCR using Gabor Filter and HMM
• - Family of two-dimensional Gabor functions was
  proposed by Daugman to model the spatial
  summation properties (of the receptive fields) of
  simple cells in the visual cortex.
• - Provides an excellent feature set that is
  robust against noise and degradations
• - Highly experimental. Expected to minimize
  error rate of character segmentation.
• - Mimics Automatic speech recognition

12
Feature Extraction A look at gabor filter
extraction for a 0 degree filter
13
Basic approach

• Segment lines from image using pixel row scan
• Extract line width
• Extract stroke width using zero crossings
• Extract features using gabor filter by moving in
  x direction with jumpsize stroke width.
• At each x(i), make an observation sequence by
  moving in y-direction.
• Find the best hmm which matches with the
  sequence.
• Repeat the process.

14
Tibetan OCR

• It is proposed to develop an OCR for the Tibetan
  script.
• The methods developed in the previous modules
  will also be applied for the Tibetan script.
• The primary motivation is that it will provide a
  useful test-bed for checking the other
  technologies.

15
Time targeted deliverables

• Annotation tool / Ground Truth Generator
• 0 6 months Basic design to be completed
  including determination of format for the XML
  document.
• 7 15 months Implementation of editor
  (multilingual) and I/O interface for images in
  various formats.
• 16 24 months Development of module for
  producing noise / degradation of input images.
• 24 - 30 months Development of the tool for
  performing error analysis of the output of OCR.
• 30 36 months Testing, debugging and
  documentation of the entire module including
  installer.

16
Time targeted deliverables

• Development of general framework
• 0 4 months Exploring various technologies
  available for creating a plug and play
  environment.
• 4 8 months Design of the system.
• 8 16 months Development of the system.
• 17 20 months Testing, debugging and
  documentation of the system.
• 21 36 months Implementation of various
  techniques / algorithms that are used in creating
  OCRs together with their implementation in the
  system.

17
Time targeted deliverables

• Visual component recognizer engine using
• Gabor filter
• 0 6 months Development of sub-modules for
  extracting extracting line width, stroke width
  and other relevant parameters
• 6 18 months Determination of the Gabor
  function parameters as functions of the document
  level parameters extracted earlier including
  extent of scaling of images.
• 19 30 months Development of at least three
  classifiers for performing the classification
• 31 36 months Testing, debugging and
  documentation of the complete tool including
  installer.

18
Time targeted deliverables

• Line component recognizer engine using HMM
• 0 6 months Design of the basic system
• 7 12 months Developing a module for performing
  HMM training and classification
• 13 18 months Developing a training set having
  different font styles and sizes in Devanagari
• 19 30 months Training the HMM and fine-tuning
  the parameters for optimal performance
  (recognition and time)
• 31 36 months Testing, debugging and
  documentation of the system including installer.

19
Time targeted deliverables

• Tibetan OCR
• 0 6 months Analysis of Tibetan script
• 7 12 months Development of training set for
  Tibetan script
• 12 30 months Design and development of OCR for
  Tibetan script
• 31 36 months Testing, debugging and
  documentation of the system including installer.