|
In this thesis, we provide a solution not only for segmenting
characters in mixed Chinese/English printed documents but also
segmenting touching characters. we divide the segmented
character into five classes: Chinese, alphanumeric, punctuation
mark, Chinese touching character and alphanumeric touching
character. Firstly, for confirmation of Chinese characters
formed by single rectangle, we calculate the horizontal and
vertical stroke crossing counts for character complexity
analysis, if its omplexity is high, and it fits to the square
style and the size of Chinese characters, then we confirm it is
a Chinese character. Secondly, for confirmation of Chinese
characters formed by two rectangles, if it also fits to the
square style and the size of Chinese characters, then we give
the largest rectangle to alphanumeric OCR, if the rectangle is
rejected, the joined character must be a Chinese character.
Otherwise, the joined rectangles are not treated as a Chinese
character. We determine the attribute of touching characters by
contextual relationship and information provided by self. The
attribute of characters could make a judgement for correction
the erroneous attributes. For the segmentation of touching
characters, we adopt the strategy with two stages, the first
stage is to search all possible cutting positons, the second
stage is to search the proper cutting positions. Experiments to
12 documents, there are 5536 Chinese characters and 1672
alphanumeric characters among 7378 characters. The result is
that 14 Chinese characters are segmented erroneously and
alphanumeric characters are 35. The correct rate is 99.7% and
97.9%. The overall correct rate is 99.3%.
|