Building Multi Script OCR for Brahmi Scripts: Selection of Efficient Features

1
Building Multi Script OCR for Brahmi Scripts
Selection of Efficient Features

• Md. Abul Hasnat

Center for Research on Bangla Language Processing
(CRBLP), Department of Computer Science and
Engineering, BRAC University, Dhaka, Bangladesh.
2
Brahmi Script Analysis

• Features of the graphemes of the characters
• Baseline / Matraa
• Vertical bar
• Curvatures

Baseline / Matraa
Curvatures
Vertical Bar
3
Brahmi Script Analysis

• Vowels have dependent and independent form.
• Vowel change shape when followed by consonant.

• Consonant followed by consonant creates new shape.

4
Bangla Script
5
Other Scripts
Devanagari
Gurmukhi
6
Other Scripts
Tibetan
Sharda
7
Outline

• Feature extraction for OCR.
• Classification
• Analysis of feature extraction approaches
• Conclusion

8
What is Feature Extraction?

• Devijver and Kittle define feature extraction as
  the problem of extracting from the raw data the
  information which most relevant for
  classification purposes, in the sense of
  minimizing the within-class pattern variability
  while enhancing the between-class pattern
  variability.
• Image features are unique characteristics that
  can represent a specific image.
• Meaningful, detectable parts of the image.
• Overcome the vulnerabilities of template
  matching.
• reduce the computation cost.

9
Feature Extraction in OCR

• The selection of image features and corresponding
  extraction methods is probably the most important
  step in achieving high performance for an OCR
  system.

• The preprocessing stage aims to make the image be
  suitable for different feature extraction
  algorithms.

10
Feature Extraction in OCR

• Properties of image features
• Robust to transformations
• Robust to noise
• Feature extraction efficiency
• Feature matching efficiency

• Issues in feature extraction
• Invariants
• features remains unchanged when a particular
  transformation is applied.
• Reconstruction
• can be reconstructed from the extracted features.

11
Features and Classifiers

• Different feature type may need different type of
  classifiers.
• Graph description - structural or syntactic
  classifiers.
• Discrete features - decision trees.
• Real valued features - statistical classifier.

12
Types of Features

• Feature extraction methods are based on three
  types of features
• Statistical
• Projections and profiles
• Crossing and distance
• Zoning
• Structural
• Nodal features
• Stroke analysis
• Global transformation and shape based
• Unitary image transform
• Shape (boundary region based)

13
Statistical Features (Projection Histograms)

• Introduced in 1956 in hardware OCR system.
• Today, this technique is mostly used for
• Segmenting characters, words, and text lines
• Detect if an input image is rotated.

Vertical Projection
Horizontal Projection
14
Statistical Features (Profile)

• Count distance between the bounding box and the
  edge of a character image.
• Used to extract the contour of the character
  image.

15
Statistical Features (Crossing)

• Count the number of transitions from background
  to foreground pixels.

V 2
H 3
Figure crossing
16
Statistical Features (Distance)

• Count the distance of the first Image pixel
  detected from upper and lower boundaries.

U 6
L 5
R 6
B 7
Figure crossing and Distance
17
Limitations (Projection, Profile, Crossing
Distance)

• Scale dependent.
• Sensitive to rotation.
• Sensitive to the variability in writing style.
• Important information about the character shape
  seems to be lost.

18
Statistical Features (Zoning)

• Divide the character image (matrix) into certain
  number of zones (sub-matrix).
• Apply computation on each zone separately.
• The goal of zoning is to obtain the local
  characteristics instead of global
  characteristics.
• Calculation over each zone
• Percentage of black pixels.
• Weight of each zone.
• Evaluate the extent to which sub-matrix shape
  matches any direction. (Used for MLP based
  classifier)

19
Structural Features (Zoning)
9 X 7
Two rows overlap
One row overlap
Weight Matrix ( 9 X 7 )
One row overlap
Two rows overlap
One row overlap
60 Degree Path
Figure Zones of a 32 X 24 image
20
Statistical Features (Zoning)

• Observations
• Additional features needed to improve the
  classifier performance.
• Overlapping between zones to enhance the
  reliability of the features.

21
Structural Features (Nodal features)
Figure Nodes extracted from a character image
22
Global Transformation (Unitary Image Transform)

• Reduction in the number of features.
• Preserving most of the information.
• Pixels are ordered by their variance, and the
  pixels with the highest variance are used as
  features.
• Reconstruction ability.

• Limitations
• Not rotation invariant
• Input image have to be exactly the same size
  (Scaling and resampling is necessary if the size
  can vary)

23
Global Transformation (Unitary Image Transform)

• Several transformation methods
• Karhunen-Loeve (KL) computationally demanding
• Fourier recommended by andrew
• Hadamard (or Walsh) -- recommended by andrew
• Haar transform
• Cosine computationally reasonable, better in
  terms of image compression
• Sine
• Slant Transform

We applied Discrete Cosine Transform with Hidden
Markov Model (HMM) as a classifier.
24
Global Transformation (Discrete Cosine Transform)
Table Reconstruction result of different
variance difference
0.7
Table Number of features for different variance
difference
25
Shape Based

• Features are invariants to translation, scale,
  rotation, blur and noise.
• Most commonly used image features, where shape
  representation is the most important issue.
• Classified into two categories
• boundary-based invariants
• region-based invariants.

26
Shape (Boundary Based)

• Explore only the contour information
• Two techniques
• Chain code
• Fourier descriptors
• Cannot capture the interior content of the shape.
• Reconstruction ability.

• Limitations
• cannot deal with disjoint shapes

• Decisions
• Not appropriate for us.

27
Shape (Region Based)

• All of the pixels of the image are taken into
  account to represent the shape.
• Can capture some of the global properties.
• Popular region-based methods
• Hus seven moment invariants
• Zernike moments
• Can also be employed to describe disjoint shapes.
  
• Reconstruction ability.

28
Region Based(Hus moment invariants)

• Seven moments
• Hus invariants have the properties of being
  invariant
• image translation
• scaling
• rotation

Table Hus seven moments

• Compute the higher order of Hus moment
  invariants is quite complex.

29
Region Based(Zernike moments)

• Allow independent moment invariants to be
  constructed easily to an arbitrarily high order.
• Concept of orthogonal moments to recover the
  image.
• Invariants to
• Rotation
• Normalized Zernike moments, Invariants to
• Translation
• Scale
• Rotation

30
Region Based(Zernike moments)
Table Number of features for different order of
Zernike moment
Table Reconstruction result for different order
of Zernike moment
31
Features s of the existing open-source OCRs

• OCROPUS Features used by the system currently
  include
• Gradients
• Singular points of the skeleton
• Presence of holes and
• Unary-coded geometric information
• Location relative to the baseline and
• Original aspect ratio and skew prior to skew
  correction.

32
Features s of the existing open-source OCRs

• Tesseract
• Feature used by tesseract includes
• Segments of the polynomial approximation.
• Direction of the outline
• For test character features are three
  dimensional
• x position
• y - position
• angle
• For training character features are three
  dimensional
• x position
• y - position
• angle
• length

33
Features s of the existing open-source OCRs

• GOCR
• Feature used by GOCR includes
• size
• skew
• presence of serifs

34
Conclusion

• Unique features can extract from the similar
  Brahmi scripts.
• Zonal features are useful as secondary features.
• Nodal features are useful if properly extracted.
• Moments are useful primary features.
• Hus seven features.
• Zernike features up to 40 order.

35
References
1 D. Trier, A.K. Jain, and T. Taxt, "Feature
extraction methods for character recognition - a
survey," Pattern Recognition, vol. 29, no. 4, pp.
641-662, Apr. 1996. 2 Tinku Acharya and Ajoy
K. Ray, Image Processing Principles and
Applications. New JerseyJohn Wiley Sons,
2005. 3 Qing Chen, "EVALUATION OF OCR
ALGORITHMS FOR IMAGES WITH DIFFERENT SPATIAL
RESOLUTIONS AND NOISES", Graduate Thesis Report,
School of Information Technology and Engineering,
Faculty of Engineering, University of
Ottawa. 4 Peter Burrow, "Arabic Handwriting
Recognition", Graduate Thesis Report, School of
Informatics, University of Edinburgh. 5 Md.
Abul Hasnat, S. M. Murtoza Habib, and Mumit Khan,
Segmentation free Bangla OCR using HMM Training
and Recognition, Proc. of 1st DCCA2007, Irbid,
Jordan, 2007. 6 R. Kapoor, D. Bagai and T.S.
Kamal, Representation and Extraction of Nodal
Features of DevNagri Letters, Proceedings of the
3rd Indian Conference on Computer Vision,
Graphics and Image Processing. 7 Jan Flusser ,
Moment Invariants in Image Analysis,
TRANSACTIONS ON ENGINEERING, COMPUTING AND
TECHNOLOGY, V11, Feb. 2006, ISSN 1305-5313 9
Liu Maofu, He Yanxiang and Ye Bin, "Image
Zernike moments shape feature evaluation based on
image reconstruction", Geo-spatial Information
Science, Volume 10, Issue 3 , May 31, 2007. 10
http//www.micro.dibe.unige.it/Research/OCR.htm
11 www.iit.demokritos.gr/IIT_SS/Presentations/Off
-Line20Handwritten20OCR.ppt 12
http//tesseract-ocr.repairfaq.org/tess_glossary.h
tml
36
Questions
37
Thank You