Parallel-Hierarchical Neural Network for 3D Object Recognition

Noriaki Sato, Masafumi Hagiwara

Research output: Contribution to journalArticle

Abstract

In this paper, the authors propose a parallel-hierarchical neural network that can recognize multiple 3D objects from 2D projection images. The proposed network focuses on a parallel-hierarchical structure and memory-based recognition assistance, which are characteristics of the excellent vision systems that living organisms have, and refers to the neocognitron, which models the parallel-hierarchical structure of vision systems. The amount of calculations is reduced by deleting cells having a low degree of importance based on competition between cells and detecting features that differ for each cell. The network not only can recognize an object, but can also estimate its orientation at the same time. The memory-based recognition assistance is modeled by performing iterative processing during which the input image approaches a learning image based on the orientation estimation result. Weights are determined by sequentially presenting learning images, and no teaching signal is necessary. This is done in a short time since it is not an iterative learning method. The network performance was evaluated by using five objects from COIL-100. Images that were obtained by photographing each object after it was rotated 60° at a time around the vertical axis were used for the learning images. The experiments checked the recognition rates for various 2D projection images that were obtained from the 3D objects. These results verified the effectiveness of the proposed neural network as a 3D object recognition technique.

Original languageEnglish
Pages (from-to)1-12
Number of pages12
JournalSystems and Computers in Japan
Volume35
Issue number1
DOIs
Publication statusPublished - 2004 Jan
Externally publishedYes

Fingerprint

3D Object Recognition
Hierarchical Networks
Object recognition
Neural Networks
Neural networks
Data storage equipment
Network performance
Teaching
Vision System
Hierarchical Structure
Processing
Cell
Projection
Experiments
Network Performance
Vertical
Object
Necessary
Learning

Keywords

  • 3D object recognition
  • Affine transformation recognition
  • Occlusion recognition
  • Vision-based recognition model

ASJC Scopus subject areas

  • Hardware and Architecture
  • Information Systems
  • Theoretical Computer Science
  • Computational Theory and Mathematics

Cite this

Parallel-Hierarchical Neural Network for 3D Object Recognition. / Sato, Noriaki; Hagiwara, Masafumi.

In: Systems and Computers in Japan, Vol. 35, No. 1, 01.2004, p. 1-12.

Research output: Contribution to journalArticle

@article{0703ed7d51774b299803bbbabcaa7b8f,
title = "Parallel-Hierarchical Neural Network for 3D Object Recognition",
abstract = "In this paper, the authors propose a parallel-hierarchical neural network that can recognize multiple 3D objects from 2D projection images. The proposed network focuses on a parallel-hierarchical structure and memory-based recognition assistance, which are characteristics of the excellent vision systems that living organisms have, and refers to the neocognitron, which models the parallel-hierarchical structure of vision systems. The amount of calculations is reduced by deleting cells having a low degree of importance based on competition between cells and detecting features that differ for each cell. The network not only can recognize an object, but can also estimate its orientation at the same time. The memory-based recognition assistance is modeled by performing iterative processing during which the input image approaches a learning image based on the orientation estimation result. Weights are determined by sequentially presenting learning images, and no teaching signal is necessary. This is done in a short time since it is not an iterative learning method. The network performance was evaluated by using five objects from COIL-100. Images that were obtained by photographing each object after it was rotated 60° at a time around the vertical axis were used for the learning images. The experiments checked the recognition rates for various 2D projection images that were obtained from the 3D objects. These results verified the effectiveness of the proposed neural network as a 3D object recognition technique.",
keywords = "3D object recognition, Affine transformation recognition, Occlusion recognition, Vision-based recognition model",
author = "Noriaki Sato and Masafumi Hagiwara",
year = "2004",
month = "1",
doi = "10.1002/scj.10579",
language = "English",
volume = "35",
pages = "1--12",
journal = "Systems and Computers in Japan",
issn = "0882-1666",
publisher = "John Wiley and Sons Inc.",
number = "1",

}

TY - JOUR

T1 - Parallel-Hierarchical Neural Network for 3D Object Recognition

AU - Sato, Noriaki

AU - Hagiwara, Masafumi

PY - 2004/1

Y1 - 2004/1

N2 - In this paper, the authors propose a parallel-hierarchical neural network that can recognize multiple 3D objects from 2D projection images. The proposed network focuses on a parallel-hierarchical structure and memory-based recognition assistance, which are characteristics of the excellent vision systems that living organisms have, and refers to the neocognitron, which models the parallel-hierarchical structure of vision systems. The amount of calculations is reduced by deleting cells having a low degree of importance based on competition between cells and detecting features that differ for each cell. The network not only can recognize an object, but can also estimate its orientation at the same time. The memory-based recognition assistance is modeled by performing iterative processing during which the input image approaches a learning image based on the orientation estimation result. Weights are determined by sequentially presenting learning images, and no teaching signal is necessary. This is done in a short time since it is not an iterative learning method. The network performance was evaluated by using five objects from COIL-100. Images that were obtained by photographing each object after it was rotated 60° at a time around the vertical axis were used for the learning images. The experiments checked the recognition rates for various 2D projection images that were obtained from the 3D objects. These results verified the effectiveness of the proposed neural network as a 3D object recognition technique.

AB - In this paper, the authors propose a parallel-hierarchical neural network that can recognize multiple 3D objects from 2D projection images. The proposed network focuses on a parallel-hierarchical structure and memory-based recognition assistance, which are characteristics of the excellent vision systems that living organisms have, and refers to the neocognitron, which models the parallel-hierarchical structure of vision systems. The amount of calculations is reduced by deleting cells having a low degree of importance based on competition between cells and detecting features that differ for each cell. The network not only can recognize an object, but can also estimate its orientation at the same time. The memory-based recognition assistance is modeled by performing iterative processing during which the input image approaches a learning image based on the orientation estimation result. Weights are determined by sequentially presenting learning images, and no teaching signal is necessary. This is done in a short time since it is not an iterative learning method. The network performance was evaluated by using five objects from COIL-100. Images that were obtained by photographing each object after it was rotated 60° at a time around the vertical axis were used for the learning images. The experiments checked the recognition rates for various 2D projection images that were obtained from the 3D objects. These results verified the effectiveness of the proposed neural network as a 3D object recognition technique.

KW - 3D object recognition

KW - Affine transformation recognition

KW - Occlusion recognition

KW - Vision-based recognition model

UR - http://www.scopus.com/inward/record.url?scp=0346055412&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0346055412&partnerID=8YFLogxK

U2 - 10.1002/scj.10579

DO - 10.1002/scj.10579

M3 - Article

AN - SCOPUS:0346055412

VL - 35

SP - 1

EP - 12

JO - Systems and Computers in Japan

JF - Systems and Computers in Japan

SN - 0882-1666

IS - 1

ER -