Sparse feature convolutional neural network with cluster max extraction for fast object classification

Sung Hee Kim, Dong Sung Pae, Tae Koo Kang, Dong W. Kim, Myo Taeg Lim

Research output: Contribution to journalArticle

Abstract

We propose the Sparse Feature Convolutional Neural Network (SFCNN) to reduce the volume of convolutional neural networks (CNNs). Despite the superior classification performance of CNNs, their enormous network volume requires high computational cost and long processing time, making real-time applications such as online-training difficult. We propose an advanced network that reduces the volume of conventional CNNs by producing a region-based sparse feature map. To produce the sparse feature map, two complementary region-based value extraction methods, cluster max extraction and local value extraction, are proposed. Cluster max is selected as the main function based on experimental results. To evaluate SFCNN, we conduct an experiment with two conventional CNNs. The network trains 59 times faster and tests 81 times faster than the VGG network, with a 1.2% loss of accuracy in multi-class classification using the Caltech101 dataset. In vehicle classification using the GTI Vehicle Image Database, the network trains 88 times faster and tests 94 times faster than the conventional CNNs, with a 0.1% loss of accuracy.

Original languageEnglish
Pages (from-to)2468-2478
Number of pages11
JournalJournal of Electrical Engineering and Technology
Volume13
Issue number6
DOIs
Publication statusPublished - 2018 Nov 1

Fingerprint

Neural networks
Processing
Costs
Experiments

Keywords

  • Classification
  • Deep learning
  • Object recognition
  • Online-training control

ASJC Scopus subject areas

  • Electrical and Electronic Engineering

Cite this

Sparse feature convolutional neural network with cluster max extraction for fast object classification. / Kim, Sung Hee; Pae, Dong Sung; Kang, Tae Koo; Kim, Dong W.; Lim, Myo Taeg.

In: Journal of Electrical Engineering and Technology, Vol. 13, No. 6, 01.11.2018, p. 2468-2478.

Research output: Contribution to journalArticle

@article{f16bbf3e9d0d4ed0ad54026b93dc420e,
title = "Sparse feature convolutional neural network with cluster max extraction for fast object classification",
abstract = "We propose the Sparse Feature Convolutional Neural Network (SFCNN) to reduce the volume of convolutional neural networks (CNNs). Despite the superior classification performance of CNNs, their enormous network volume requires high computational cost and long processing time, making real-time applications such as online-training difficult. We propose an advanced network that reduces the volume of conventional CNNs by producing a region-based sparse feature map. To produce the sparse feature map, two complementary region-based value extraction methods, cluster max extraction and local value extraction, are proposed. Cluster max is selected as the main function based on experimental results. To evaluate SFCNN, we conduct an experiment with two conventional CNNs. The network trains 59 times faster and tests 81 times faster than the VGG network, with a 1.2{\%} loss of accuracy in multi-class classification using the Caltech101 dataset. In vehicle classification using the GTI Vehicle Image Database, the network trains 88 times faster and tests 94 times faster than the conventional CNNs, with a 0.1{\%} loss of accuracy.",
keywords = "Classification, Deep learning, Object recognition, Online-training control",
author = "Kim, {Sung Hee} and Pae, {Dong Sung} and Kang, {Tae Koo} and Kim, {Dong W.} and Lim, {Myo Taeg}",
year = "2018",
month = "11",
day = "1",
doi = "10.5370/JEET.2018.13.6.2468",
language = "English",
volume = "13",
pages = "2468--2478",
journal = "Journal of Electrical Engineering and Technology",
issn = "1975-0102",
publisher = "Korean Institute of Electrical Engineers",
number = "6",

}

TY - JOUR

T1 - Sparse feature convolutional neural network with cluster max extraction for fast object classification

AU - Kim, Sung Hee

AU - Pae, Dong Sung

AU - Kang, Tae Koo

AU - Kim, Dong W.

AU - Lim, Myo Taeg

PY - 2018/11/1

Y1 - 2018/11/1

N2 - We propose the Sparse Feature Convolutional Neural Network (SFCNN) to reduce the volume of convolutional neural networks (CNNs). Despite the superior classification performance of CNNs, their enormous network volume requires high computational cost and long processing time, making real-time applications such as online-training difficult. We propose an advanced network that reduces the volume of conventional CNNs by producing a region-based sparse feature map. To produce the sparse feature map, two complementary region-based value extraction methods, cluster max extraction and local value extraction, are proposed. Cluster max is selected as the main function based on experimental results. To evaluate SFCNN, we conduct an experiment with two conventional CNNs. The network trains 59 times faster and tests 81 times faster than the VGG network, with a 1.2% loss of accuracy in multi-class classification using the Caltech101 dataset. In vehicle classification using the GTI Vehicle Image Database, the network trains 88 times faster and tests 94 times faster than the conventional CNNs, with a 0.1% loss of accuracy.

AB - We propose the Sparse Feature Convolutional Neural Network (SFCNN) to reduce the volume of convolutional neural networks (CNNs). Despite the superior classification performance of CNNs, their enormous network volume requires high computational cost and long processing time, making real-time applications such as online-training difficult. We propose an advanced network that reduces the volume of conventional CNNs by producing a region-based sparse feature map. To produce the sparse feature map, two complementary region-based value extraction methods, cluster max extraction and local value extraction, are proposed. Cluster max is selected as the main function based on experimental results. To evaluate SFCNN, we conduct an experiment with two conventional CNNs. The network trains 59 times faster and tests 81 times faster than the VGG network, with a 1.2% loss of accuracy in multi-class classification using the Caltech101 dataset. In vehicle classification using the GTI Vehicle Image Database, the network trains 88 times faster and tests 94 times faster than the conventional CNNs, with a 0.1% loss of accuracy.

KW - Classification

KW - Deep learning

KW - Object recognition

KW - Online-training control

UR - http://www.scopus.com/inward/record.url?scp=85055627228&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85055627228&partnerID=8YFLogxK

U2 - 10.5370/JEET.2018.13.6.2468

DO - 10.5370/JEET.2018.13.6.2468

M3 - Article

VL - 13

SP - 2468

EP - 2478

JO - Journal of Electrical Engineering and Technology

JF - Journal of Electrical Engineering and Technology

SN - 1975-0102

IS - 6

ER -