Á¤º¸°úÇÐȸ ³í¹®Áö D : µ¥ÀÌŸº£À̽º
ÇѱÛÁ¦¸ñ(Korean Title) |
SOSiM: ÇüÅ Ư¡ ±â¼úÀÚ¸¦ »ç¿ëÇÑ ÇüÅ ±â¹Ý °´Ã¼ À¯»ç¼º ¸ÅĪ |
¿µ¹®Á¦¸ñ(English Title) |
SOSiM: Shape-based Object Similarity Matching using Shape Feature Descriptors |
ÀúÀÚ(Author) |
³ëÃæÈ£
À̼®·æ
Á¤Áø¿Ï
±è»óÈñ
±è´öȯ
Chungho Noh
Seoklyong Lee
Chinwan Chung
Sanghee Kim
Deokhwan Kim
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 36 NO. 02 PP. 0073 ~ 0083 (2009. 04) |
Çѱ۳»¿ë (Korean Abstract) |
º» ³í¹®¿¡¼´Â ¿µ»ó ³»ÀÇ °´Ã¼ÀÇ ÇüÅÂ(shape)¿¡ ±â¹ÝÇÑ °´Ã¼ À¯»ç¼º ¸ÅĪ(matching) ¹æ¹ýÀ» Á¦¾ÈÇÑ´Ù. Á¦¾ÈÇÑ ¹æ¹ý¿¡¼´Â °´Ã¼ÀÇ À±°û¼±(edge)¿¡¼ Á¡µé(edge points)À» ÃßÃâÇÏ°í, ÃßÃâµÈ Á¡µéÀÇ À§Ä¡ °ü°è¸¦ ³ªÅ¸³»±â À§ÇÏ¿© °¢ Á¡À» ±âÁØÀ¸·Î ·Î±× ¿øÇü È÷½ºÅä±×·¥(log polar histogram)À» »ý¼ºÇÏ¿´´Ù. °´Ã¼ÀÇ À±°ûÀ» µû¶ó°¡¸ç °¢ Á¡¿¡ ´ëÇÑ ¿øÇü È÷½ºÅä±×·¥À» ¼øÂ÷ÀûÀ¸·Î ºñ±³ÇÔÀ¸·Î½á °´Ã¼°£ÀÇ ¸ÅĪÀÌ ÀÌ·ç¾îÁö¸ç, µ¥ÀÌŸº£À̽º·ÎºÎÅÍ À¯»çÇÑ °´Ã¼¸¦ °Ë»öÇϱâ À§ÇÏ¿© »ç¿ëÇÑ ¸ÅĪ ¹æ½ÄÀº ³Î¸® ¾Ë·ÁÁø k-NN (nearest neighbor) ÁúÀÇ ¹æ½ÄÀ» »ç¿ëÇÏ¿´´Ù. Á¦¾ÈÇÑ ¹æ¹ýÀ» °ËÁõÇϱâ À§ÇÏ¿© ±âÁ¸ÀÇ ÇüÅ ¹®¸Æ ±â¹ý (Shape Context method)°ú Á¦¾ÈÇÑ ¹æ¹ýÀ» ºñ±³ÇÏ¿´À¸¸ç, °´Ã¼ À¯»ç¼º ¸ÅĪ ½ÇÇè¿¡¼ k=5ÀÏ ¶§ ±âÁ¸ ¹æ¹ýÀÇ Á¤È®µµ°¡ 0.37, Á¦¾ÈÇÑ ¹æ¹ýÀÌ 0.75-0.90À̸ç, k=10ÀÏ ¶§ ±âÁ¸ ¹æ¹ýÀÌ 0.31, Á¦¾ÈÇÑ ¹æ¹ýÀÌ 0.61-0.80·Î¼ ±âÁ¸ÀÇ ¹æ¹ý¿¡ ºñÇØ Á¤È®ÇÑ ¸ÅĪ °á°ú¸¦ º¸¿© ÁÖ¾ú´Ù. ¶ÇÇÑ ¿µ»óÀÇ È¸Àü º¯Çü ½ÇÇè¿¡¼ ±âÁ¸ ¹æ¹ýÀÇ Á¤È®µµ°¡ 0.30, Á¦¾ÈÇÑ ¹æ¹ýÀÌ 0.69·Î¼ ±âÁ¸ ¹æ¹ýº¸´Ù ȸÀü º¯Çü¿¡ °ÀÎÇÑ(robust) Ư¼ºÀ» °¡ÁüÀ» °üÂûÇÒ ¼ö ÀÖ¾ú´Ù. |
¿µ¹®³»¿ë (English Abstract) |
In this paper we propose an object similarity matching method based on shape characteristics of an object in an image. The proposed method extracts edge points from edges of objects and generates a log polar histogram with respect to each edge point to represent the relative placement of extracted points. It performs the matching in such a way that it compares polar histograms of two edge points sequentially along with edges of objects, and uses a well-known k-NN(nearest neighbor) approach to retrieve similar objects from a database. To verify the proposed method, we¡¯ve compared it to an existing Shape-Context method. Experimental results reveal that our method is more accurate in object matching than the existing method, showing that when k=5, the precision of our method is 0.75-0.90 while that of the existing one is 0.37, and when k=10, the precision of our method is 0.61-0.80 while that of the existing one is 0.31. In the experiment of rotational transformation, our method is also more robust compared to the existing one, showing that the precision of our method is 0.69 while that of the existing one is 0.30.
|
Å°¿öµå(Keyword) |
°´Ã¼ À¯»ç¼º ¸ÅĪ
¿µ»ó °Ë»ö
°´Ã¼ ÀνÄ
ÇüÅ Ư¡ ±â¼úÀÚ
k-ÃÖ±ÙÁ¢ ÁúÀÇ
object similarity matching
image retrieval
object recognition
shape feature descriptor
k-nearest neighbor query
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|