Orawan Jungthanawong. A high performance system for printed Thai character recognition with a stroke structure approach. Master's Degree(Computer Science). Mahidol University. : Mahidol University, 2007.
A high performance system for printed Thai character recognition with a stroke structure approach
Abstract:
In this thesis, a highly accurate recognition system for Thai printed characters is
proposed. High recognition accuracy leads to high efficiency in further processing,
for example, text to speech or machine translation, etc.
To achieve a highly accurate recognition system, the causes of erroneous
recognition have to be solved. From the investigation, it was found that one of major
causes is that the number of characters to be recognized is too high. The higher the
number is, the larger the search space will become. In order to reduce the search
space, this thesis proposes a 2-stage process. In the first stage, the characters are
grouped into 12 clusters according to the shape, which are consonants, special
symbols and Thai digits such as clusters of {ข,ฃ,ช,ซ}, {บ,ป,ษ}, etc. Similarly, 8
clusters of 21 characters which are vowels and tone marks can be categorized, for
instance clusters of {โ, ไ, ใ}, { -, -, -, -}, etc. In the second stage, the character is
recognized with the domain within each class. At this stage, features distinct within
the class and indistinct across the class can be used. Thus, the accuracy is improved.
In the segmentation process, characters of an upper-vowel level and a tonal level are
combined into a single-character image.
The proposed system is tested with 100,000 Thai characters. The type tested
fonts were Browallia New, Cordia New and FreesiaUPC .The sizes of the font were
14, 16 and 18. Both normal and bold styled fonts were tested. The proposed system
achieved 100% recognition rate in the test under the controlled environment. The
average analysis time and the average image and program storage in hard-disk were
363 ms per a character and 5755.5 MB per 1 A4 paper page, respectively.