Connected Component Level Method Identification in Automatic Titleboard Indexing

1
Connected Component Level Method Identification
in Automatic Titleboard Indexing

• Samuel James Pinson
• Mark Pinson
• Dr. William Barrett
• Computer Science Department
• Brigham Young University

2
Connected Component LevelMethod Identification

• Distinguishing between machine print and
  handwriting

3
Microfilm Titleboards

• Type
• Location
• Time
• Acquisition

4
Microfilm Titleboards
5
Problem Statement

• To make genealogical microfilm more accessible by
  automatically building a searchable index over
  titleboards
• Preprocess titleboards
• Distinguish between machine print and handwriting
  (Method identification)
• Recognize machine print and handwriting (OCR)
• Build searchable index

6
Method Identification

• Discriminator Between Handwritten and
  Machine-Printed Characters Umeda et al. 1990
• U.S. Patent 4,910,787

Handwriting
Slanted
Horizontal
Vertical
7
Method Identification

• Separating Handwritten Material from Machine
  Printed Text Using Hidden Markov Models Guo et
  al. 2001

50
50
55
45
38
62
33
67
21
79
23
77
Machine Print
Handwriting
8
Method Identification

• Machine Printed Text and Handwriting
  Identification in Noisy Document Images Zheng et
  al. 2004

Noise
Noise
Handwriting
Noise
Machine Print
Handwriting
Handwriting
Machine Print
Machine Print
9
Method Identification

• Eigenfaces Turk and Pentland 1991

N
2
N
0
71
250
68
210
44
128
53
N
10
Method Identification
11
Method Identification
12
Method Identification

• The Use of Eigenpictures for Optical Character
  Recognition Muller and Herbst 1998

13
Method Identification
14
Method Identification

• Determining a local distance threshold via radial
  density

Global Target Precision
98
Local Precision
100
99
92
15
Index Construction
Archivo
de
la
Parroquia
INDEX
BAUTISMO
115
1877
1878
Red.
12-1
Fecha
9-27-1960
16
Querying
ltbautismogt AND lt1960gt
INDEX
17
Results
18
Results
19
Future Work

• Robust preprocessing and segmentation
• Incorporate lexical, font, and style context.
• Metadata about indexed terms script, language,
  meaning
• Specialize the set of representative machine
  print connected components

20
Conclusions

• Connected component level method identification
• Progress towards automatic titleboard indexing