Online Handwritten Shape Recognition Using Segmental Hidden Markov Models

1
Online Handwritten Shape RecognitionUsing
Segmental Hidden Markov Models

• Presented by Yueping Qian

2
Problem Definition

• PEN-BASED interfaces have been recently
  popularized in many fields
• electronic paper, tablet PC notebooks,
  electronic pen, Personal Digital Assistants
• Many recognition systems are target at specific
  application and can not be used to other tasks

3
Base shape
4
Terms

• Stroke-Level Representation
• description of a character as a sequence of sub
  strokes Each sub-stroke is represented by three
  features a base shape chosen from a finite
  dictionary and its size and position

5
Illustration of the Stroke-Level Representation
6
General architecture of the recognition system.
7
Pseudo Code

• An input character signal is first normalized by
  setting the height and width of its bounding box
  to unity, we can represent this character as a
  sequence of observations OO1.Ot
• Then O are segmented into
• a sequence of sub-stroke rn
• The transformation from O to base shape is
  performed using the angle sequence
• hn is a base shape from BS, dn is its relative
  duration, and pn is its position

8
Pseudo Code

• 1. Preprocessing and, segment the character into
  sub-stokes using SLR model (decoding), a sequence
  of base shapes are produced.
• substroke sequence that maximizes the probability
  of o
• 2. building an allograph model
• 3. divisive clustering algorithm is performed to
  select representative models

9
Building an allograph model

• A unique state sequence qK associated with R
• The final probability for the allograph model is
  computed using a Viterbi approximation

10
Experiments result

• 1. Result of Learning Different Types of
  Characters

11
Experiments result

• 2. The result of Learning with Few Training
  Samples

12
Experiments result

• omparisons of the best performing systems for
  digit recognition with others like SVMDTW, DTW,
  SN-Tuples, SlrHmm, bn, and HMM

13
Examples
14
Examples
15
Thanks!