May 14, 2004
Comparison Between Multi - Vector Feature Space Method and Earth Mover's Distance Method In Similarity Searches of Images
IPSJ SIG Notes
- ,
- ,
- ,
- ,
- ,
- Volume
- 2004
- Number
- 45
- First page
- 83
- Last page
- 90
- Language
- Japanese
- Publishing type
- Publisher
- Information Processing Society of Japan (IPSJ)
There are some cases where dissimilar images are judged to be similar in using quadratic-form distance, a representative one in similarity searches of images. To address this problem, we proposed a multi-vector feature space method based on pseudo-Euclidean space and an oblique basis (MVPO). In this method, an image is represented by a solid consisting of multiple vectors; each vector corresponds to each feature. We also proposed D-distance as a distance between solids. A representative method similar to ours is Earth Mover's Distance (EMD), which is said to experimentally outperforms others including quadratic-form distance in precision. We show that EMD can be formalized as a distance in MVPO, and a reason why it outperforms quadratic-form distance. We also mention the difference between D-distance and EMD.
- Link information
-
- CiNii Articles
- http://ci.nii.ac.jp/naid/110002911530
- CiNii Books
- http://ci.nii.ac.jp/ncid/AN10114171
- URL
- http://id.nii.ac.jp/1001/00040249/
- ID information
-
- CiNii Articles ID : 110002911530
- CiNii Books ID : AN10114171