Design of an Automated CNN Composition Scheme with Lightweight Convolution for Space-Limited Applications

Feng Hao Yeh, Ding Chau Wang, Pi Wei Chen, Pei Ju Li, Wei Han Chen, Pei Hsuan Yu, Chao Chun Chen

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

The emergence of the CNN network has enabled many networks for image object recognition, object segmentation, etc., and has brought amazing results to image processing tasks, including MaskRCNN [4] and YOLO [8]. These networks can achieve comparable performance by stacking Convolutional Layers, as layers go deeper, the performance is improved as well. Although deeper convolution layers make the performance of the entire network better, the huge parameters of the networks makes it difficult to implement the network on embedded systems with constrained hardware resources. Therefore, if these networks are to be run on devices with resource constraint, the structure of the network must be lightweight. Usually the most prevalent way to reduce the weight of the network is to modify the network structure, but the design of the network structure has its own philosophy. Any changes to the structure of the network will compromise the performance of the network. We propose a method to automatically substitute the Convolution architecture in the network without changing the network architecture, thereby reducing the parameter of the network while ensuring the performance of the network.

Original languageEnglish
Title of host publicationIntelligent Information and Database Systems - 15th Asian Conference, ACIIDS 2023, Proceedings
EditorsNgoc Thanh Nguyen, Bogumiła Hnatkowska, Siridech Boonsang, Hamido Fujita, Tzung-Pei Hong, Kitsuchart Pasupa, Ali Selamat
PublisherSpringer Science and Business Media Deutschland GmbH
Pages436-446
Number of pages11
ISBN (Print)9789819958337
DOIs
Publication statusPublished - 2023
EventProceedings of the 15th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2023 - Phuket, Thailand
Duration: 2023 Jul 242023 Jul 26

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13995 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

ConferenceProceedings of the 15th Asian Conference on Intelligent Information and Database Systems, ACIIDS 2023
Country/TerritoryThailand
CityPhuket
Period23-07-2423-07-26

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'Design of an Automated CNN Composition Scheme with Lightweight Convolution for Space-Limited Applications'. Together they form a unique fingerprint.

Cite this