Problem
Tool
Harder problem
Experiments
Result
Optical Character Recognition using Bayesian
Networks
Ioannis Klasinas
iklasinas@telecom.tuc.gr
July 11, 2007
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
Problem
Tool
Harder problem
Experiments
Result
Problem
Letter Recognition Using Holland-Style Adptive Classifiers, Peter
W. Frey, David J. Slate
English capital letters
20000 instances (bitmap fonts)
45x45 pixel bitmap
Images distorted (linear magnification, aspect radio,
horizontal/vertical wrap)
16 features extracted
82.7% accuracy
Others 93,6% (Statlog ALLOC80)
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
Problem
Tool
Harder problem
Experiments
Result
Weka
Weka (http://www.cs.waikato.ac.nz/ml/weka/)
Various classification methods
Used Bayes networks
87.5% accuracy, 4 parents per node
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
Problem
Tool
Harder problem
Experiments
Result
Digit OCR
Scanned handwritten digits
16x16 grayscale bitmaps
9200 instances
Threshold to convert to b/w
Extracted features
Normalized as above
NRR-1:94.5%, Bayes:38.2%
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
Problem
Tool
Harder problem
Experiments
Result
Experiments
Experimented with
1 threshold
2 max parents number
Best result for threshold=0.2, max parents=16
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
Problem
Tool
Harder problem
Experiments
Result
Threshold
Figure: Bitmaps, for threshold -0.5/0/0.5
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks
84
th=-0.5
th=-0.4
83 th=-0.3
th=-0.2
th=-0.1
th=0
82 th=0.1
th=0.2
th=0.3
81 th=0.4
80
79
78
77
76
75
74
0 2 4 6 8 10 12 14 16 18
Figure: Results
Problem
Tool
Harder problem
Experiments
Result
Discussion
Handwritten OCR tough problem
Weka unpredictable
Bayesian networks inferior to other approaches for this
problem
More appropriate features needed
Ioannis Klasinas iklasinas@telecom.tuc.gr Optical Character Recognition using Bayesian Networks