Support Vector Machines for Handwritten Numerical String Recognition

1
Support Vector Machinesfor Handwritten Numerical
String Recognition
9th International Workshop on Frontiers in
Handwriting Recognition October, 26-29, 2005
Tokyo, JAPAN

• Luiz S. Oliveira and Robert Sabourin

Pontifical Catholical University of Parana
(PUCPR), Curitiba, BRAZIL Ecole de Technologie
Supérieure (ETS), Montreal, CANADA
2
Introduction

• In the last years, SVMs have gained a lot of
  attention.
• ML and PR communities.
• Sucessfully applied to several different areas
  such as
• Face Verification/Recognition, Image Retrieval,
  Speech Recogntion, Text Categorization,
  Handwriting Recogntion.

3
Introduction

• Handwriting Recognition
• Handwritten digits (MNIST) have been used as
  benchmark by ML and PR communities.
• ML raw image since their goal is to assess the
  technique.
• PR preoccupied in achieving performance on a
  given database
• Feature Extraction.

4
Introduction

• Performance of SVM-based classifiers

5
Motivation

• Evaluate SVMs in the context of strings of
  digits.
• Problems of under- and over-segmentation.
• Baseline system
• MLP-based classifiers and two verifiers
  specialized to deal with the efects of
  segmentation.
• MLP is not robust to deal with outliers.

6
Baseline System
PAMI, 24(11), 2002.
7
Heuristic Over-Segmentation
Among all possible combinations, we have these
pieces.
8
SVM

• Binary classifier for separable classes.
• The Kernel Trick
• Makes it suitable for non-linear decision
  surfaces.
• Linear algorithm to operate in a higher
  dimensionality space.
• More than two classes
• Pairwise, one-against-others.

9
Pairwise
Winner
In this strategy, the number of classifiers we
have to train is q(q-1)/2, e.g., 45 for q10.
10
One-Against-Others
11
Estimating Probabilities

• SVM produces an uncalibrated value that is not a
  probability.
• There are several situations where would be
  useful to have a classifier producing a
  posteriori probability.
• In our case, the Baseline system is based on a
  probabilistic framework.

12
Platts Method

• Fits the scores produced by the SVMs into a
  logistic function
• The parameters A and B are found by minimizing
  the negative log likelihood of the training data.

13
Implementation
Replace the general-purpose classifier and remove
the verifiers.
SVMs
14
Implementation

• Train models
• 10 SVMs (One-against-others).
• Gaussian Kernels.
• Parameters estimated through a grid search
• C 215, 214, ..., 2-2
• ? 24, 23, ..., 2-10
• Performance of (C,?) was evaluated on a
  validation set.

15
Implementation

• Train parameters of the sigmoid.
• Platts algorithm.
• Same training set used to train the SVMs.
• Database
• NIST SD19
• 195000, 28000, and 60089 for training,
  validation, and testing respectively.

16
Results

• Isolated digits
• Strings of digits
• 12800 strings ranging from 2 to 10 digits.

17
Results
NV Without Verifiers V With Verifiers
18
Results
19
Conclusion

• SVM applied to recognize string of digits.
• SVMs are better suited than MLPs for outliers
  resulting from improper segmentation.
• More time consuming.

20
Thanks!!