Eigenspace Method by Autoassociative Networks for Object Recognition

Takamasa Yokoi, Wataru Oyama, Tetsushi Wakabayashi, Fumitaka Kimura

Research output: Contribution to journalArticle

Abstract

This paper studies on a new eignespace method which employs autoassociative networks for object recognition. Five layered autoassociative network is available to obtain a manifold on the minimum square error hypersurface which approximates a distribution of learning sample. Recognition experiments were performed to show that the manifold of rotating object is obtained by learning and the objects, such as a mouse and a stapler, are correctly recognized by the autoassociative networks. It is also shown that the accuracy of approximating closed manifold and the accuracy of recognition are improved by emploing multiple autoassociative networks each of which is trained by a partition of the learning sample.The property and the advantage of the five layered autoassociative network are demonstrated by a comparative study with the nearest neighbor method and the eigenspace method.

Original languageEnglish
Pages (from-to)95-103
Number of pages9
JournalLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume3138
Publication statusPublished - Dec 1 2004
Externally publishedYes

Fingerprint

Eigenspace
Object recognition
Object Recognition
Learning
Experiments
Nearest Neighbor Method
Comparative Study
Hypersurface
Mouse
Rotating
Partition
Recognition (Psychology)
Closed
Experiment
Object

All Science Journal Classification (ASJC) codes

  • Computer Science(all)
  • Biochemistry, Genetics and Molecular Biology(all)
  • Theoretical Computer Science

Cite this

Eigenspace Method by Autoassociative Networks for Object Recognition. / Yokoi, Takamasa; Oyama, Wataru; Wakabayashi, Tetsushi; Kimura, Fumitaka.

In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Vol. 3138, 01.12.2004, p. 95-103.

Research output: Contribution to journalArticle

@article{f45e0017ce624d028ec6777829190427,
title = "Eigenspace Method by Autoassociative Networks for Object Recognition",
abstract = "This paper studies on a new eignespace method which employs autoassociative networks for object recognition. Five layered autoassociative network is available to obtain a manifold on the minimum square error hypersurface which approximates a distribution of learning sample. Recognition experiments were performed to show that the manifold of rotating object is obtained by learning and the objects, such as a mouse and a stapler, are correctly recognized by the autoassociative networks. It is also shown that the accuracy of approximating closed manifold and the accuracy of recognition are improved by emploing multiple autoassociative networks each of which is trained by a partition of the learning sample.The property and the advantage of the five layered autoassociative network are demonstrated by a comparative study with the nearest neighbor method and the eigenspace method.",
author = "Takamasa Yokoi and Wataru Oyama and Tetsushi Wakabayashi and Fumitaka Kimura",
year = "2004",
month = "12",
day = "1",
language = "English",
volume = "3138",
pages = "95--103",
journal = "Lecture Notes in Computer Science",
issn = "0302-9743",
publisher = "Springer Verlag",

}

TY - JOUR

T1 - Eigenspace Method by Autoassociative Networks for Object Recognition

AU - Yokoi, Takamasa

AU - Oyama, Wataru

AU - Wakabayashi, Tetsushi

AU - Kimura, Fumitaka

PY - 2004/12/1

Y1 - 2004/12/1

N2 - This paper studies on a new eignespace method which employs autoassociative networks for object recognition. Five layered autoassociative network is available to obtain a manifold on the minimum square error hypersurface which approximates a distribution of learning sample. Recognition experiments were performed to show that the manifold of rotating object is obtained by learning and the objects, such as a mouse and a stapler, are correctly recognized by the autoassociative networks. It is also shown that the accuracy of approximating closed manifold and the accuracy of recognition are improved by emploing multiple autoassociative networks each of which is trained by a partition of the learning sample.The property and the advantage of the five layered autoassociative network are demonstrated by a comparative study with the nearest neighbor method and the eigenspace method.

AB - This paper studies on a new eignespace method which employs autoassociative networks for object recognition. Five layered autoassociative network is available to obtain a manifold on the minimum square error hypersurface which approximates a distribution of learning sample. Recognition experiments were performed to show that the manifold of rotating object is obtained by learning and the objects, such as a mouse and a stapler, are correctly recognized by the autoassociative networks. It is also shown that the accuracy of approximating closed manifold and the accuracy of recognition are improved by emploing multiple autoassociative networks each of which is trained by a partition of the learning sample.The property and the advantage of the five layered autoassociative network are demonstrated by a comparative study with the nearest neighbor method and the eigenspace method.

UR - http://www.scopus.com/inward/record.url?scp=35048892515&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=35048892515&partnerID=8YFLogxK

M3 - Article

AN - SCOPUS:35048892515

VL - 3138

SP - 95

EP - 103

JO - Lecture Notes in Computer Science

JF - Lecture Notes in Computer Science

SN - 0302-9743

ER -