A Study of an Analysis System Based on Deep Learning for Mirror-like Texture
碩士 === 國立雲林科技大學 === 電子工程系 === 107 === The surface detail texture of metal may be unclear owing to the high specular reflections. Therefore, it was not always easy to distinguish the level of polishing in frosted metal. Today, the level of polishing of frosted metal was mainly classified by human sou...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | zh-TW |
Published: |
2019
|
Online Access: | http://ndltd.ncl.edu.tw/handle/q654u2 |
id |
ndltd-TW-107YUNT0393039 |
---|---|
record_format |
oai_dc |
spelling |
ndltd-TW-107YUNT03930392019-10-17T05:52:20Z http://ndltd.ncl.edu.tw/handle/q654u2 A Study of an Analysis System Based on Deep Learning for Mirror-like Texture 基於深度學習的鏡面紋理分析檢測系統之研究 HO,PENG-LUN 何朋倫 碩士 國立雲林科技大學 電子工程系 107 The surface detail texture of metal may be unclear owing to the high specular reflections. Therefore, it was not always easy to distinguish the level of polishing in frosted metal. Today, the level of polishing of frosted metal was mainly classified by human sources. But the production efficiency was affected by inconsistent measurement results due to differences in people's judgment. In the experiment, we used the ring light source to be the light source of detection. To achieve the uniform light source, using the white card to calibrate the image which captured by camera. Obtaining the different frosted characteristics of steel tubes with intensity distribution of ring light source that analyzed by three dimensional image. At first, we transformed colorful image into grayscale image for simplifying the complicated computation. Second, using Gaussian low pass filter to remove the noise of grayscale image. Third, we obtained the frosted characteristics of measurement object by canny edge detection. Fourth, using the region of interest to get the characteristics. Finally, we used local binary pattern to analyze the different texture. We could classify the frosted kind of different sandpaper after the analysis of texture. In the research, we trained the image with convolutional neural networks. Convolutional neural networks was an effective method when we wanted to sort the characteristics of two-dimensional image. We analyzed the difference of different texture by using supervised learning in the convolutional neural networks. The supervised learning could achieve 86.5% precision after the textured classification in the P100 and P1000 sandpaper. Finally, it could approach the industry standard. LIN,CHING-HUANG 林慶煌 2019 學位論文 ; thesis 81 zh-TW |
collection |
NDLTD |
language |
zh-TW |
format |
Others
|
sources |
NDLTD |
description |
碩士 === 國立雲林科技大學 === 電子工程系 === 107 === The surface detail texture of metal may be unclear owing to the high specular reflections. Therefore, it was not always easy to distinguish the level of polishing in frosted metal. Today, the level of polishing of frosted metal was mainly classified by human sources. But the production efficiency was affected by inconsistent measurement results due to differences in people's judgment. In the experiment, we used the ring light source to be the light source of detection. To achieve the uniform light source, using the white card to calibrate the image which captured by camera. Obtaining the different frosted characteristics of steel tubes with intensity distribution of ring light source that analyzed by three dimensional image.
At first, we transformed colorful image into grayscale image for simplifying the complicated computation. Second, using Gaussian low pass filter to remove the noise of grayscale image. Third, we obtained the frosted characteristics of measurement object by canny edge detection. Fourth, using the region of interest to get the characteristics. Finally, we used local binary pattern to analyze the different texture. We could classify the frosted kind of different sandpaper after the analysis of texture.
In the research, we trained the image with convolutional neural networks. Convolutional neural networks was an effective method when we wanted to sort the characteristics of two-dimensional image. We analyzed the difference of different texture by using supervised learning in the convolutional neural networks. The supervised learning could achieve 86.5% precision after the textured classification in the P100 and P1000 sandpaper. Finally, it could approach the industry standard.
|
author2 |
LIN,CHING-HUANG |
author_facet |
LIN,CHING-HUANG HO,PENG-LUN 何朋倫 |
author |
HO,PENG-LUN 何朋倫 |
spellingShingle |
HO,PENG-LUN 何朋倫 A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
author_sort |
HO,PENG-LUN |
title |
A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
title_short |
A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
title_full |
A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
title_fullStr |
A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
title_full_unstemmed |
A Study of an Analysis System Based on Deep Learning for Mirror-like Texture |
title_sort |
study of an analysis system based on deep learning for mirror-like texture |
publishDate |
2019 |
url |
http://ndltd.ncl.edu.tw/handle/q654u2 |
work_keys_str_mv |
AT hopenglun astudyofananalysissystembasedondeeplearningformirrorliketexture AT hépénglún astudyofananalysissystembasedondeeplearningformirrorliketexture AT hopenglun jīyúshēndùxuéxídejìngmiànwénlǐfēnxījiǎncèxìtǒngzhīyánjiū AT hépénglún jīyúshēndùxuéxídejìngmiànwénlǐfēnxījiǎncèxìtǒngzhīyánjiū AT hopenglun studyofananalysissystembasedondeeplearningformirrorliketexture AT hépénglún studyofananalysissystembasedondeeplearningformirrorliketexture |
_version_ |
1719269774283767808 |