A Modular System to Recognize Numerical Amounts on Brazilian Bank Cheques

1
A Modular System to RecognizeNumerical Amounts
on Brazilian Bank Cheques

• 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
System Overview

• Segmentation-based recognition.
• Explicit segmentation.
• Integration of all modules is done through a
  probabilistic model.
• Problem to overcome
• To distinguish, at the recognition stage,
  isolated (correctly segmented) characters from
  over and under segmentation.
• Recognition and verification approach.

3
Over and Under-segmentationProblems.
Misclassification caused by over-segmentation (a
and b) and under-segmentation (c)
4
Modular System
Grey boxes AD modules. White boxes TD modules.
5
Component Detection and Segmentation

• Component Detection
• It operates in three steps connected component
  analysis, delimiter detection and grouping.
• Segmentation
• Relationship among complementary structural
  features contour, profile and skeleton
  IWFHR00.
• Segmentation graphs.

6
Features and General-purpose Recognizer

• General-purpose recognizer.
• Mixture of concavity and contour features.
• e10 and e3 132 components.
• e13 18 components (13 outputs 4 structural
  features 1 contextual feature).
• Databases 11 400, 2 000 and 4 000 (training,
  validation and testing).
• Performance 99.2, 99.0 and 98.9.

7
Verifier

• In order to overcome over- and under-segmentation
  problems, we have proposed the following
  verifier
• MLP with 3 classes isolated, over-segmented and
  under-segmented characters.
• Features.
• Multi-level concavity analysis.
• Profile distances.
• Databases.
• 40 500, 4 000 and 4 000 (training, validation and
  testing).
• 99.02 on test set.

8
New Feature Set
42 components from MCA 6 components from
profile distances 48 components.
9
Interaction Between GPR and Verifier
10
Global Hypothesis and Post-processor

• Hypothesis generation.
• Modified Viterbi algorithm.
• Post-processor.
• Deterministic automaton.

11
Experimental Results

• Experiments on numerical amounts.
• 503 images (about 9 characters per image).

Recognition Rates (zero-rejection level)
12
Experimental Results

• NIST SD19.
• Database.
• Isolated digits 195 000, 28 000 and 60 000
  (training, validation and testing).
• Performance on isolated digits (zero-rejection
  level) 99.66, 99.55 and 99.13.
• Verifier 40 500, 4 000 and 4 000 (training,
  validation and testing).
• Performance on test set 98.90.
• Database of strings 12 800 images (hsf_7 series).

13
Experimental Results
Recognition Rates (zero-rejection level)
14
Experimental Results
Recognition rates on NIST database reported by
other authors.

• ICDAR, (2) New Results.
• Results achieved without knowledge about the
  number of digits in the strings.

15
Future Works

• Future Works.
• Optmization of the classifiers.
• General-purpose recognizer.
• Verifier.
• Optimization of the system.
• Ensemble of classifiers.