ﮐﺎﻫﺶ ﻓﻀﺎي ﺟﺴﺘﺠﻮ در ﺑﺎزﺷﻨﺎﺳﯽ زﯾﺮ ﮐﻠﻤﺎت ﺗﺎﯾﭙﯽ ﻓﺎرﺳﯽ ﺑﺎ اﺳﺘﻔﺎده از وﯾﮋﮔﯽ ﻫﺎي زﯾﺴﺖ ﺳﻨﺠﻪ ﻣﯿﻨﻮﺷﯿﺎ

Search Space Reduction in Farsi Machine-printed Sub-words Recognition by using Biometric Minutia Features

ﭼﮑﯿﺪه: ﺑﺎ ﺗﻮﺟﻪ ﺑﻪ ﮔﺴﺘﺮده ﺑﻮدن زﯾﺮﮐﻠﻤﺎت ﺗﺎﯾﭗ ﺷﺪه ﻓﺎرﺳﯽ، ﯾﺎﻓﺘﻦ ﯾﮏ زﯾﺮﮐﻠﻤﻪ و ﺑﻪ ﺗﺒﻊ آن ﯾﮏ ﮐﻠﻤﻪ در ﯾﮏ ﻣﺘﻦ ﭼﺎپ ﺷﺪه ﮐﺎر ﺑﺴﯿﺎر زﻣﺎن ﺑﺮي ﺧﻮاﻫﺪ ﺑﻮد. در اﯾﻦ ﻣﻘﺎﻟﻪ، روﺷﯽ ﻣﺒﺘﻨﯽ ﺑﺮ ﻧﻘﺎط زﯾﺴﺖ ﺳﻨﺠﻪ ﻣﯿﻨﻮﺷﯿﺎ اراﺋﻪ ﺷﺪه اﺳﺖ ﮐﻪ ﻓﻀﺎي ﺟﺴﺘﺠﻮي زﯾﺮﮐﻠﻤﺎت ﺗﺎﯾﭗ ﺷﺪه ﻓﺎرﺳﯽ را ﺑﻪ ﺻﻮرت ﻗﺎﺑﻞ ﺗﻮﺟﻬﯽ ﮐﺎﻫﺶ ﻣﯽ دﻫﺪ. ﻟﺬا ﺗﻌﺪاد ﻧﻘﺎط و ﻣﺨﺘﺼﺎت ﻣﯿﻨﻮﺷﯿﺎي اﻧﺸﻌﺎﺑﯽ و اﻧﺘﻬﺎﯾﯽ ﮐﻪ دو وﯾﮋﮔﯽ ﻣﻄﺮح در ﺣﻮزه زﯾﺴﺖ ﺳﻨﺠﻪ ﻣﯽ ﺑﺎﺷﻨﺪ، ﺑﻌﻨﻮان وﯾﮋﮔﯽ ﻫﺎﯾﯽ ﺑﺮاي ﮐﺎﻫﺶ ﻓﻀﺎي ﺟﺴﺘﺠﻮ در ﻗﺎﻟﺐ ﯾﮏ روش دو ﻣﺮﺣﻠﻪ اي اﺳﺘﻔﺎده ﺷﺪه اﻧﺪ. در ﮔﺎم ﻧﺨﺴﺖ ﻧﻘﺎط ﻣﯿﻨﻮﺷﯿﺎ از ﺗﺼﻮﯾﺮ زﯾﺮﮐﻠﻤﻪ اﺳﺘﺨﺮاج ﺷﺪه و در ﭼﻬﺎرﺧﻮﺷﻪ ﮐﻪ از ﻟﺤﺎظ ﺗﻌﺪاد ﻧﻘﺎط ﺑﻪ ﯾﮑﺪﯾﮕﺮ ﻧﺰدﯾﮏ ﻫﺴﺘﻨﺪ دﺳﺘﻪ ﺑﻨﺪي ﻣﯽ ﺷﻮﻧﺪ، ﺑﻪ اﯾﻦ ﺗﺮﺗﯿﺐ ﻓﻀﺎي ﺟﺴﺘﺠﻮ ﺗﻘﺮﯾﺒﺎً ﻧﺼﻒ ﺧﻮاﻫﺪ ﺷﺪ. در ﮔﺎم دوم ﺑﺎ اﯾﺠﺎد ﯾﮏ ﻣﺨﺰن از ﻓﻮاﺻﻞ اوﻟﯿﻦ ﺗﺎ آﺧﺮﯾﻦ ﻧﻘﻄﻪ اﻧﺘﻬﺎﯾﯽ ﺑﺮاي ﻫﺮ زﯾﺮﮐﻠﻤﻪ در ﻫﺮ ﺧﻮﺷﻪ و ﺗﻄﺒﯿﻖ ﻓﺎﺻﻠﻪ ﻣﺬﮐﻮر در ﺗﺼﻮﯾﺮ آزﻣﺎﯾﺸﯽ ﺑﺎ ﻣﺨﺰن، ﻓﻀﺎي ﺟﺴﺘﺠﻮ ﺑﻪ ﻣﻘﺪار ﻗﺎﺑﻞ ﺗﻮﺟﻬﯽ ﮐﺎﻫﺶ ﻣﯽ ﯾﺎﺑﺪ. ﻧﺘﺎﯾﺞ ﺑﺪﺳﺖ آﻣﺪه از اﻋﻤﺎل روش ﭘﯿﺸﻨﻬﺎدي ﺑﺮ روي ﺗﺼﺎوﯾﺮ زﯾﺮﮐﻠﻤﻪ ﻣﻮﺟﻮد در ﭘﺎﯾﮕﺎه داده ﻧﺸﺎن ﻣﯽ دﻫﺪ، ﻓﻀﺎي ﺟﺴﺘﺠﻮ از 12700 زﯾﺮﮐﻠﻤﻪ ﺑﻄﻮر ﻣﺘﻮﺳﻂ ﺣﺪود 98/8 درﺻﺪ، ﺑﺎ دﻗﺖ ﺗﻘﺮﯾﺒﯽ ﺑﯿﺸﺘﺮ از 98 درﺻﺪ ﮐﺎﻫﺶ ﯾﺎﻓﺘﻪ اﺳﺖ.

Due to the wide range of Persian (Farsi) machine-printed sub-words, finding a sub-word and consequently a word in a machine-printed text will be very time consuming. In this paper, a biometric minutia based method is presented, which significantly reduces the search space of Farsi sub-words. Therefore, the number of points and their coordinates of bifurcated and ending minutia points, which are two well-known features in the field of biometrics, have been used as features to reduce the search space in the form of a two-step method. In the first step, the minutia points are extracted from the sub-word image and categorized into four clusters that are close and similar as a viewpoint of the number of minutia points. Therefore, search space will be halved by this step. In the second step, by creating a repository of the distances between the first and the last end points for each sub-word in each cluster and matching the same distance of experimental image with the repository, the search space is significantly reduced. Obtained results show that the search space has been reduced from 12,700 sub-words to about 500 sub-keywords with an accuracy of approximately 90%.