Extraction of character strings printed on products

1
Extraction of character strings printed on
products

• M. Okamoto, H. Kawajiri and K. Soma
• Department of Information Engineering,
• Shinshu University

2
Purpose and Past Research

• Purpose
• Development of OCR for characters printed on
  products
• Past Research
• Automatic inspection in manufacturing
• Vehicle license number reader
• Character recognition in scene images

3
What are characters

• One of most difficult problems in OCR
• Some features of character strings
• Aspect ratio, size, pitch, etc. of enclosing
  rectangles,
• Color information
• Score of classifier for connected component

4
Object images

• Characters are printed with the similar color and
  size in a character string ( word, text line
  etc.)
• Characters are printed in horizontal or vertical
  direction with a small skew
• Characters are printed on flat or round surfaces
• No limitation on background colors and patterns
• No limitation on lighting conditions

5
Overiew of our algorithm
6
Luv uniform color space
7
A test image
8
White pixels
Black pixels
Cluster 1
9
Black pixels
White pixels
Cluster 2
10
(No Transcript)
11
(No Transcript)
12
(No Transcript)
13
Neighboring rectangles
14
Alignment condition of rectangles
15
The extracted characters
16
Input image
17
Cluster 1
18
Cluster 2
19
Cluster 3
20
Cluster 4
21
Extracted characters
22
(No Transcript)
23
(No Transcript)
24
Conclusions and Further research

• More than 70 of characters are extracted except
  for heavily blurred or merged ones
• Color information and heuristic knowlage on
  character strings are useful
• Further research
• Improvement of the clustering algorithm
• Implementation of a OCR with our extraction method