Character Recognition Experimental Design

1
Character Recognition Experimental Design

• Michael S. Spiegel
• DePauw University

2
Techniques of the Past
3
Generalized Character Recognition

• Alphabet is created (2 common ways)
• Character samples from human writers
• Expert intelligence about the letterforms
• A character is drawn
• Character compared against alphabet
• CR system returns the letter represented by the
  most equivalent alphabet character

4
Types of CR

• On-line
• eg, Palm OS
• Alphabet Independent
• General pattern recognizer
• User Dependent

• Off-line
• eg, OCR
• Alphabet Dependent
• eg, Sanskrit or English recognition
• User Independent

5
Testing a CR System

• Alphabet and test data obtained from
• Private collection study
• Collection style geared toward specifics of the
  CR system (advantage)
• Provides apple and orange comparisons between
  system results (disadvantage)
• Existing data corpus
• At the mercy of the collection style
  (disadvantage)
• Affords reliable comparisons between system
  results (advantage)

6
Data Corpuses
7
Testing a CR System

• Collections often approximate either
• Language letter frequencies (alphabet and
  language dependent)
• Realistic writing conditions (email or text)
• Difficult to have both in one collection

8
Testing a CR System

• Data set is often divided into
• Training set / alphabet
• Test set
• Each letter/case must exist at least twice in a
  data set
• Frequencies 2 count minimum usually manifest in
  a single static training set!

9
Problem Statement

• Design a collection and experiment methodology
  that provides for
• User independent (and therefore user dependent)
  systems
• Alphabet and language independence
• Character and traditional letter frequency
  analysis

10
Our Collection Method

• Requires 30 instances of each character
• Upper- and lower-case
• Training data is not included in test data
• More instances needed

• Testing alphabet sizes 1 to 3
• 30 3 33 instances of each character of each
  case from each participant needed for statistical
  significance
• 33 26 2 1716 characters per user!

11
To Apply Letter Frequency

• 33 instances of each character
• q 33 e 1878
• Too much writing!!
• 16,820 total characters
• Extremely difficult to collect
• Write for an entire day
• Conduct multiple testing sessions per participant
• (Oxford Dictionary)

12
Letter Frequency Weighting

• Sample error 27.4
• Frequency weighted error 10.1

13
Varied Size of Alphabet

• Chose multiple random alphabets of each size per
  user
• Alphabets of size
• 1 3326 3.0e39
• 2 (3332)26 4.1e78
• 3 (333231)26 2.5e117
• Chose 900 examples to reach 99 confidence of the
  users mean accuracy within .086

14
Example In Action

• Previous Graffiti study found a user to have
  97.5 accuracy with
• 1 alphabet containing 3 of each letter, and one
  test run
• Same user with our new method

15
Conclusions

• On-line
• User-dependent
• Alphabet-independent
• Novel approach to satisfy these things
• 60 individuals tested
• 33 instances per character/case
• Alphabets size 1 to 3

16
Questions?