KCC 2021
Current Result Document :
ÇѱÛÁ¦¸ñ(Korean Title) |
ÀǷ῵»ó ºÐ·ù¸¦ À§ÇÑ °Å¸®±â¹Ý ¸ÞÆ®¸¯ ¼Õ½ÇÀÇ ºñ±³ |
¿µ¹®Á¦¸ñ(English Title) |
Comparison of Distance-Based Metric Losses for Medical Image Classification |
ÀúÀÚ(Author) |
±â¹Ì¸£ ·¯¸¸
ÀÌ»ó¿õ
Raman Ghimire
Sang-Woong Lee
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 48 NO. 01 PP. 0657 ~ 0658 (2021. 06) |
Çѱ۳»¿ë (Korean Abstract) |
|
¿µ¹®³»¿ë (English Abstract) |
Distance metric learning has been used as an effective strategy to learn intrinsic pattern of the image datasets across various computer vision domain. In this paper, we compare 3 different types of loss functions: pair-based loss (Triplet loss, Soft-Triple loss) and proxy-based loss (Proxy-NCA, Proxy-Anchor) in a multi-class classification task. We have used resnet-50 as network architecture and our result shows that Proxy-Anchor loss achieved 93.44% mean average precision on average compared to the Triplet loss, SoftTriple Loss and Proxy-NCA loss which achieved 92.53%, 93.63% and 91.98% respectively. |
Å°¿öµå(Keyword) |
Medical Image Classification
Triplet Loss
Soft-Triple Loss
Proxy-NCA Loss
Proxy-Anchor Loss
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|