Á¤º¸°úÇÐȸ ³í¹®Áö A : ½Ã½ºÅÛ ¹× ÀÌ·Ð
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
¿µ»ó Áß¿äµµ Á¤º¸¿¡ ±â¹ÝÇÑ ½ºÅ¸ÀÏ ÀüÀÌ ±â¹ý |
¿µ¹®Á¦¸ñ(English Title) |
Style Transfer Technique based on Image Saliency |
ÀúÀÚ(Author) |
ÀÌȣâ
À±°æÇö
Hochang Lee
Kyunghyun Yoon
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 40 NO. 02 PP. 0111 ~ 0116 (2013. 04) |
Çѱ۳»¿ë (Korean Abstract) |
½ºÅ¸ÀÏ ÀüÀÌ ±â¹ýÀº ÂüÁ¶ ¿µ»ó¿¡ ³ªÅ¸³ª´Â Áú°¨/»ö»ó°ú °°Àº ½ºÅ¸ÀÏÀ» °á°ú ¿µ»ó¿¡ ¹Ý¿µÇÏ´Â ±â¹ýÀÌ´Ù. ±âÁ¸ÀÇ ½ºÅ¸ÀÏ ÀüÀÌ ±â¹ý ¿¬±¸µéÀº »ö»óÀ̳ª Áú°¨ Áß ÇϳªÀÇ Æ¯Â¡¿¡ ÃÊÁ¡À» ¸ÂÃß¾î ÁøÇàµÇ¾úÀ¸¸ç, µÎ °¡Áö Ư¡À» µ¿½Ã¿¡ ¹Ý¿µÇÑ ¿¬±¸´Â È°¹ßÈ÷ ¿¬±¸µÇÁö ¾Ê¾Ò´Ù. ¶ÇÇÑ ½ºÅ¸ÀÏ ÀüÀ̸¦ À§ÇØ Èĺ¸ Çȼ¿ÁýÇÕÀ» ±¸ÇÒ ¶§ ´Ü¼øÈ÷ ÀÓÀÇÀÇ Çȼ¿µé·Î ±¸¼ºÇϱ⠶§¹®¿¡ ¶§¶§·Î ¾È ÁÁÀº °á°ú ¿µ»óÀ» »ý¼ºÇÏ¿´´Ù. º» ¿¬±¸¿¡¼´Â ÂüÁ¶ È¸È ÀÛÇ°ÀÇ »ö»ó°ú Áú°¨ Á¤º¸¸¦ µ¿½Ã¿¡ ¹Ý¿µÇÏ´Â ¾Ë°í¸®ÁòÀ» Á¦¾ÈÇÑ´Ù. À̶§ ÂüÁ¶ ¿µ»óÀÇ ½ºÅ¸ÀÏÀ» È¿°úÀûÀ¸·Î ¹Ý¿µÇϱâ À§ÇØ ¿µ»óÀÇ Áß¿äµµ Á¤º¸¸¦ ÀÌ¿ëÇÑ´Ù. ¿ì¸®´Â ÂüÁ¶ ¿µ»ó°ú ´ë»ó¿µ»óÀÇ Áß¿äµµ¸¦ °í·ÁÇÏ¿©, Áß¿äµµ°¡ À¯»çÇÑ ¿µ¿ª°£¿¡ ½ºÅ¸ÀÏÀÌ ¹Ý¿µµÉ ¼ö ÀÖµµ·Ï ÇÏ¿´´Ù. ¿ì¸®ÀÇ ¾Ë°í¸®ÁòÀº »ö»ó°ú Áú°¨À» µ¿½Ã¿¡ ¹Ý¿µÇϱ⠶§¹®¿¡ ±âÁ¸ ¿¬±¸¿¡ ºñÇØ ÂüÁ¶ ¿µ»óÀÇ ´À³¦À» Àß ÀüÀÌÇÒ ¼ö ÀÖ¾ú´Ù. ¶ÇÇÑ ÀÓÀÇÀÇ Çȼ¿À» ±â¹ÝÀ¸·Î Èĺ¸ ÁýÇÕÀ» ±¸¼ºÇÏ´Â °ÍÀÌ ¾Æ´Ñ, Áß¿äµµ Á¤º¸¸¦ Ãß°¡·Î ¹Ý¿µÇϱ⠶§¹®¿¡, ¿µ¿ªº°·Î º¸´Ù ¸íÈ®ÇÏ°Ô Æ¯Â¡À» ¹Ý¿µÇÒ ¼ö ÀÖ¾ú´Ù.
|
¿µ¹®³»¿ë (English Abstract) |
Style transfer re-renders a target image with texture or color taken from parts of a source image that is matched locally to the target image. Existing style transfer techniques are largely divided into two part, color and texture, and many researches of them are proposed for expressing non-photorealistic rendering based on style transfer techniques. However, most of previous researches are focused only one part (color or texture) and don¡¯t consider both color and texture information. Additionally they construct candidate pixel set for transfer by random access approach. So they sometimes make unnatural results. In this paper, we propose a novel style transfer technique that takes into account the both color and texture of the source image. Our method uses image saliency values to construct a candidate set for transfer style of source effectively. We consider saliency of target and reference image, and then match the style to region with similar saliency value. Different from previous works, we consider both color and texture information. So we can express more similar result from source image. Also we can express more realistic effect per each region because we construct candidate set based on image saliency information.
|
Å°¿öµå(Keyword) |
½ºÅ¸ÀÏ ÀüÀÌ ±â¹ý
¿µ»ó Áß¿äµµ
Çȼ¿ ±â¹Ý ·»´õ¸µ
style transfer
image saliency
pixel-based rendering
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|