Highlevel Verification of Handwritten Numeral Strings

1
High-level Verification of Handwritten Numeral
Strings

• L.S.Oliveira, R.Sabourin, F.Bortolozzi, and
  C.Y.Suen

Pontifícia Universidade Católica do Paraná
(PUCPR) BRAZIL Ecole de Technologie Superiéure
(ETS) CANADA Centre for Pattern Recognition and
Machine Inteligence (CENPARMI) - CANADA
2
Introduction

• Significant improvements in the recognition rates
  of handwritten numeral string recognition.
• Combination of classifiers.
• Run time inefficiency.
• System complexity.
• Verification.
• Refine top few candidates in order to improve the
  recognition rate economically.

3
Kinds of Verification

• Absolute verification.
• Is it a 0.
• One-to-one verification.
• Is it a 4 or a 9.
• Verification in clustered.
• Is it a 0, 6 or a 8.

4
Levels of Verification

• High-level.
• Deal with a sub-set of the classes considered by
  the general-purpose recognizer.
• Low-level.
• Deal with meta-classes of the system such as
  characters and part of them.

5
Baseline System ICDAR01

• Recognition and verification.
• Three neural networks trained with BP.
• General-purpose recognizer 10 classes 0..9.
• Two low-level verifiers
• Over- and under-segmentation.

6
Absolute High-level Verifier

• Ten absolute verifiers (one for each numerical
  class).
• Two classes digit and noise which is composed by
  other digits.
• Concavities, moments and edge-maps.

Recognition Rates ?
7
Absolute High-level Verifier

• Supplied an improvement to naturally isolated
  digits.
• Problems with fragmented and touching digits.

8
One-to-one High-level Verifier

• Concentrate on the local difference between two
  classes.
• We have observed 39 confusions.
• Possible confusing digit pairs is 109/245.
• We can solve 75.05 and 62.76 of all errors
  focusing on top-39 and top-20 confusions
  respectively.

9
One-to-one High-level Verifier

• Top-20 confusions.

10
One-to-one High-level Verifier

• We trained 20 verifiers (top-20).
• Concavities, edge maps and histograms.

11
One-to-one High-level Verifier

• Strategy to improve this verification scheme.
• Improve the training set by including
  misrecognized samples.
• Lack of samples to improve the database.
  Considering the top-1 confusion, we have just 48
  cases.
• Including few misrecognized samples in the
  training set, probably we will introduce noise
  to our models.

12
Error Analysis

• Identify different sources of error and find why
  the high-level verification schemes achieved
  unsatisfactory results.
• Four classes of errors
• General-purpose recognizer.
• Low-level verifiers.
• Segmentation.
• Fragmentation.

13
Error Analysis of the General-purpose Recognizer

• Different confusions for different string
  lengths.
• One-to-one verifier could supply improvements for
  some string lengths but not an global
  improvement.
• Different confusions for numeral strings and
  isolated digits.
• Confusions generated by the segmentation
  algorithm such as ligatures.

14
Other Sources of Errors

• Low-level verifiers.
• Part of digits classified as an isolated digit.
• Touching digits classified as an isolated digits.
• Segmentation.
• Under-segmentation (lack of segmentation points).
• Fragmentation.
• Fragmented part grouped with the wrong neighbor.

15
Conclusion and Future Works

• High-level verification becomes interesting when
• There is a diversity of samples (confusions) to
  train the verifiers.
• For systems with weak general-purpose recognizer.
• Advantages and drawbacks (limitations) of
  high-level verification.
• Optimization of the classifiers.
• Feature selection.