Optimization of Line Segmentation Techniques for Thai Handwritten
Document
Olarik Surinta
Mahasarakham University
Thailand
Introduction
• In handwritten recognition, the line segmentation is an essential scheme.
• The occurrence of an inaccurately line segmentation will cause errors in the character segmentation.
• Most of line segmentation techniques have been based on horizontal projection profile technique.
10/21/2009 SNLP2009 2
Introduction (cont)
• The texts in most document images are aligned along horizontal lines.
• Projection profile based techniques may be one of the most successful top-down algorithms
10/21/2009 SNLP2009 3
The characteristic of Thai character
Character types CharacterConsonants ก ข ฃ ค ฅ ฆ ง จ ฉ ช ซ ฌ
ญ ฎ ฏ ฐ ฑ ฒ ณ ด ต ถ ท ธน บ ป ผ ฝ พ ฟ ภ ม ย ร ฤล ฦ ว ศ ษ ส ห ฬ อ ฮ
Vowels อ0 อะ อา อ3 อ4 อ5 อ6 อ7 อ8เอ โอ ใอ ไอ ๆ อ> อ? อ@
Tones อA อB อC อD10/21/2009 SNLP2009 4
The characteristic of Thai character (cont)
10/21/2009 SNLP2009 5
Thai sentence structure
Based on the horizontal projection profile
• The horizontal projection profile is used in dividing the text image into character line.
10/21/2009 SNLP2009 6
The line segmentation techniques in this research
• 1. The horizontal projection technique• 2. The stripe technique• 3. The comparing Thai character technique• 4. The sorting and distinguishing• (all of techniques based on horizontal projection profile)
10/21/2009 SNLP2009 7
1. The horizontal projection profile
10/21/2009 SNLP2009 8
resultHorizontal histogramImage document
2. The stripe technique
• Firstly, the stripe technique divides image into stripe (small column).
• After that, the horizontal projection profile is used to divided the text image into character lines.
10/21/2009 SNLP2009 9
2. The stripe technique (cont)
10/21/2009 SNLP2009 10
The result of strip technique for horizontal projection profile
3. Technique for comparing Thai character
• This technique takes advantage of the differences in size of characters to differentiate Thai characters between consonants and a group of small vowels and tones.
10/21/2009 SNLP2009 11
Comparing between consonant and a group of small vowel and tone
3. Technique for comparing Thai character (cont)
• First step• The groups are divided into two groups (upper
and lower zone). The higher group is then used to define the line from the image document
10/21/2009 SNLP2009 12
3. Technique for comparing Thai character (cont)
10/21/2009 SNLP2009 13
The result of first step of comparison Thai character technique
3. Technique for comparing Thai character (cont)
• Second step• consider the high value of white pixel between
the line markers and choose a new line marker
10/21/2009 SNLP2009 14
3. Technique for comparing Thai character (cont)
• this technique is complex as there are many steps to be proved.
10/21/2009 SNLP2009 15
The result of comparing Thai character technique.
4. The new technique for sorting and distinguishing
• This technique is not complicated and suitable for Thai character.
• Firstly• Use the histogram of horizontal projection profile
to sort the group of black pixels by starting with the minimum to maximum of black pixel.
10/21/2009 SNLP2009 16
4. The new technique for sorting and distinguishing (cont)
10/21/2009 SNLP2009 17
Sorting the group of black pixels.
4. The new technique for sorting and distinguishing (cont)
• Secondly• Find the maximum difference between two
groups of black pixels• The line marker is marked on the middle of the
group of black pixels when the maximum difference value is less than value of the group of black pixels
10/21/2009 SNLP2009 18
4. The new technique for sorting and distinguishing (cont)
10/21/2009 SNLP2009 19
The result of second step of sorting and distinguishing technique.
4. The new technique for sorting and distinguishing (cont)
• Finally• A new line marker is placed in the middle
between every two conjunction line markers
10/21/2009 SNLP2009 20
4. The new technique for sorting and distinguishing
10/21/2009 SNLP2009 21
Click me to play this video.
Experimental result
• Thai image documents were generated from different peoples.
• Data sets contained varieties of writing styles, and limited to only single-column Thai image documents.
10/21/2009 SNLP2009 22
Experimental result (cont)
10/21/2009 SNLP2009 23
Single-Column
Not Single-Column
Experimental result (cont)
• The line marker is used to define the character line.
• The line makers pass through the image document and do not cross the group of black pixels (line segment is completed).
10/21/2009 SNLP2009 24
Experimental result (cont)
10/21/2009 SNLP2009 25
Complete line segmentation
Experimental result (cont)
10/21/2009 SNLP2009 26
incomplete line segmentation
Experimental result (cont)
Number of lines on image documents
percentage T1 T2 T3 T4
4 46 35 92 100
5 32 26 94 99
6 26 15 88 100
7 31 24 91 96
8 21 32 90 97
9 15 11 85 95
10 18 9 88 97
11 23 12 90 94
12 7 11 88 96
Average 24.33 19.44 89.55 97.11
10/21/2009 SNLP2009 27
T1 is Horizontal projection techniqueT2 is Stripe techniqueT3 is Comparing Thai characterT4 is Sorting and distinguishing
Conclusion
• I have presented four techniques for the line segmentation of Thai language– Horizontal projection profile– Stripe– Comparing Thai character– Sorting and distinguishing
• 4 techniques based on horizontal projection profile
10/21/2009 SNLP2009 28
Conclusion (cont)
• The accuracy of the techniques are– The horizontal projection technique 24.33%– The stripe technique
19.44%
(suitable for English character and Oriya text)
– The comparing Thai character technique65.25%– The sorting and distinguishing technique97.11%
(complex, many steps to be proved)
10/21/2009 SNLP2009 29
Thank you
Question & Answer
10/21/2009 SNLP2009 30