Intelligent defect classification system based on deep learning
A fast and intelligent defect classification system for distinguishing defect features is developed in this study. Defect images obtained from an automated optical inspection instrument are first trained utilizing a deep learning approach based on the convolutional neural network. The detailed featu...
Main Authors: | , , , |
---|---|
Format: | Article |
Language: | English |
Published: |
SAGE Publishing
2018-03-01
|
Series: | Advances in Mechanical Engineering |
Online Access: | https://doi.org/10.1177/1687814018766682 |
id |
doaj-734bbe2b9f5b40b0ae47163dd7059259 |
---|---|
record_format |
Article |
spelling |
doaj-734bbe2b9f5b40b0ae47163dd70592592020-11-25T02:59:56ZengSAGE PublishingAdvances in Mechanical Engineering1687-81402018-03-011010.1177/1687814018766682Intelligent defect classification system based on deep learningRuifang Ye0Chia-Sheng Pan1Ming Chang2Qing Yu3Key Laboratory of Process Monitoring and System Optimization for Mechanical and Electrical Equipment, Huaqiao University, Xiamen, ChinaDepartment of Mechanical Engineering, Chung Yuan Christian University, Taoyuan, TaiwanDepartment of Mechanical Engineering, Chung Yuan Christian University, Taoyuan, TaiwanKey Laboratory of Process Monitoring and System Optimization for Mechanical and Electrical Equipment, Huaqiao University, Xiamen, ChinaA fast and intelligent defect classification system for distinguishing defect features is developed in this study. Defect images obtained from an automated optical inspection instrument are first trained utilizing a deep learning approach based on the convolutional neural network. The detailed features of defects, such as flaws in the inclination, size, quantity, and settlement, can then be characterized with the developed system. The obtained defect characteristics can be provided as a reference to evaluate the manufacturing process. A graphics processing unit card is used to build the parallel computing architecture for fast data computation both in the training and classification processes. The experimental results show that the defect classification for a touch panel glass surface of size 43 × 229.4 mm 2 that yields 805 million pixel data points was completed in 2 s. The classification accuracy was above 96%. Therefore, an automated and fast defect analysis catalog integrated to an optical inspection result for the evaluation and adjustment of manufacturing operations can be expected with the proposed approach.https://doi.org/10.1177/1687814018766682 |
collection |
DOAJ |
language |
English |
format |
Article |
sources |
DOAJ |
author |
Ruifang Ye Chia-Sheng Pan Ming Chang Qing Yu |
spellingShingle |
Ruifang Ye Chia-Sheng Pan Ming Chang Qing Yu Intelligent defect classification system based on deep learning Advances in Mechanical Engineering |
author_facet |
Ruifang Ye Chia-Sheng Pan Ming Chang Qing Yu |
author_sort |
Ruifang Ye |
title |
Intelligent defect classification system based on deep learning |
title_short |
Intelligent defect classification system based on deep learning |
title_full |
Intelligent defect classification system based on deep learning |
title_fullStr |
Intelligent defect classification system based on deep learning |
title_full_unstemmed |
Intelligent defect classification system based on deep learning |
title_sort |
intelligent defect classification system based on deep learning |
publisher |
SAGE Publishing |
series |
Advances in Mechanical Engineering |
issn |
1687-8140 |
publishDate |
2018-03-01 |
description |
A fast and intelligent defect classification system for distinguishing defect features is developed in this study. Defect images obtained from an automated optical inspection instrument are first trained utilizing a deep learning approach based on the convolutional neural network. The detailed features of defects, such as flaws in the inclination, size, quantity, and settlement, can then be characterized with the developed system. The obtained defect characteristics can be provided as a reference to evaluate the manufacturing process. A graphics processing unit card is used to build the parallel computing architecture for fast data computation both in the training and classification processes. The experimental results show that the defect classification for a touch panel glass surface of size 43 × 229.4 mm 2 that yields 805 million pixel data points was completed in 2 s. The classification accuracy was above 96%. Therefore, an automated and fast defect analysis catalog integrated to an optical inspection result for the evaluation and adjustment of manufacturing operations can be expected with the proposed approach. |
url |
https://doi.org/10.1177/1687814018766682 |
work_keys_str_mv |
AT ruifangye intelligentdefectclassificationsystembasedondeeplearning AT chiashengpan intelligentdefectclassificationsystembasedondeeplearning AT mingchang intelligentdefectclassificationsystembasedondeeplearning AT qingyu intelligentdefectclassificationsystembasedondeeplearning |
_version_ |
1724700188396748800 |