Á¤º¸°úÇÐȸ ³í¹®Áö B : ¼ÒÇÁÆ®¿þ¾î ¹× ÀÀ¿ë
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
½Å°æ¸Á °¡ÁöÄ¡±â¿¡ ±Ù°ÅÇÑ Æ¯Â¡ ¼±ÅÃÀÇ Çʱâü ¹®ÀÚ ÀνĿ¡ÀÇ ÀÀ¿ë |
¿µ¹®Á¦¸ñ(English Title) |
An Application of Feature Selection based on Neural Network Pruning to Handwritten Character Recognition |
ÀúÀÚ(Author) |
À±Á¾¹Î
½Å¿ä¾È
Â÷ÇüÅÂ
Á¤±Ô½Ä
Jongmin Yoon
Yoan Shin
Hyungtai Cha
Kyusik Chung
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 24 NO. 10 PP. 1043 ~ 1052 (1997. 10) |
Çѱ۳»¿ë (Korean Abstract) |
º» ³í¹®Àº ½Å°æ¸Á °¡ÁöÄ¡±â¿¡ ±Ù°ÅÇÑ Æ¯Â¡ ¼±Åà ¹æ¹ýÀÇ Çʱâü ¹®ÀÚ ÀνĿ¡ÀÇ Àû¿ëÀ» ½ÃµµÇÑ °ÍÀ¸·Î ¹× ±âÃÊ ÁÖ¿ä ½ÇÇè °á°ú¸¦ º¸¿©ÁØ´Ù. Ư¡ÀÌ ÃßÃâµÈ ÈÄ ´ÙÃþ ÆÛ¼ÁÆ®·Ð ÀνıâÀÇ ÀԷ¿¡ ÁÖ¾îÁø´Ù°í °¡Á¤ÇÏ°í, ÈÆ·ÃÀÌ ³¡³ Àνı⿡¼ Ãâ·Â°¡ÁßÄ¡°¡ ¸Å¿ì ÀÛÀº ÀÔ·Â ³ëµåµéÀ» ÈÞ¸®½ºÆ½¿¡ ±Ù°ÅÇÏ¿© »èÁ¦ÇÏ´Â »óÈ£ÀûÀÎ °¡ÁöÄ¡±â ¹æ¹ýÀ» Àû¿ëÇÑ´Ù. ÀÌ·¯ÇÑ °¡ÁöÄ¡±â¸¦ Àû¿ëÇÔÀ¸·Î½á Ư¡ °ø°£ÀÇ Â÷¿ø°¨¼Ò °á°ú¸¦ ¾ò°ÔµÈ´Ù. °¡ÁöÄ¡±â¸¦ Àû¿ëÇÑ °æ¿ì¿Í Àû¿ëÇÏÁö ¾ÊÀº °æ¿ì¿¡ ´ëÇÏ¿© ±â¿ï±â Ư¡, UDLRH ¿À¸ñ¼º Ư¡, ¸Å½¬ Ư¡, Haar Ư¡À» ÀÌ¿ëÇÏ¿© Çʱâü ¿µ¾î, ¼ýÀÚ¿Í ÇÑ±Û ÀÚ¼Ò ÀÎ½Ä ½ÇÇèÀ» ¼öÇàÇÏ¿´´Ù. ½ÇÇèÀ» ¼öÇàÇÑ °á°ú, ±â¿ï±â Ư¡°ú UDLRH ¿À¸ñ¼º Ư¡ÀÇ °æ¿ì °¡ÁöÄ¡±â¸¦ Àû¿ëÇÏÁö ¾ÊÀº °æ¿ìÀÇ Á¤Àνķü¿¡ ºñÇÏ¿© ÀνķüÀÌ ÃÖ°í 1.34% ÀúÇϵǴ ¹üÀ§¿¡¼ Ư¡ º¤ÅÍÀÇ ¾à 20%¿¡ ÇØ´çµÇ´Â ÀÔ·Â ³ëµå¸¦ »èÁ¦ÇÒ ¼ö ÀÖÀ½À» È®ÀÎÇÒ ¼ö ÀÖ¾ú´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
This paper investigates an application of feature selection based on neural network pruning to handwritten character recognition and presents some preliminary experimental results. Assuming that features are extracted and presented as the inputs of a multi-layered perceptron classifier, we apply an interactive pruning method to the trained classifier where input nodes with outgoing weights very small are pruned based on heuristic. This pruning results in a dimensionality reduction of feature space. For gradient, UDLRH concavity, mesh and Haar feature, we perform several experiments of Handwritten English Alphanumeric and Korean Alphabet recognition with/without pruning. The experimental results show that, in the case of gradient and UDLRH concavity, about 20% of feature vectors can be removed with at most 1.34% drop of the correct recognition rate of unpruned case.
|
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|