Illuminant Estimation based on Shallow Convolutional Neural Networks
碩士 === 逢甲大學 === 通訊工程學系 === 106 === White balance in imaging devices mimics the ability of color constancy in human vision. It assures an accurate color reproduction in images under a variety of illuminant conditions. The first step of white balance is to estimate the color of the illuminant. Then, t...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | zh-TW |
Published: |
2018
|
Online Access: | http://ndltd.ncl.edu.tw/handle/559apg |
id |
ndltd-TW-106FCU00650003 |
---|---|
record_format |
oai_dc |
spelling |
ndltd-TW-106FCU006500032019-05-16T00:08:07Z http://ndltd.ncl.edu.tw/handle/559apg Illuminant Estimation based on Shallow Convolutional Neural Networks 基於淺層卷積神經網路之光源顏色估測方法 WEI,YIN-JHIH 魏吟芝 碩士 逢甲大學 通訊工程學系 106 White balance in imaging devices mimics the ability of color constancy in human vision. It assures an accurate color reproduction in images under a variety of illuminant conditions. The first step of white balance is to estimate the color of the illuminant. Then, the estimated illuminant color is applied to eliminate the color cast so that the corrected image appears as taken under the canonical illuminant. The learning based approach in color constancy shows much higher performance than that of the traditional statistical approach. Especially, a convolutional neural network (CNN) enables to take feature extraction into the training process, so it can learn a complex mapping between the input and the output. The purpose of this thesis is to develop methods of estimating illuminant color using CNNs. The thesis is composed of three parts. The CNN used as a regressor [1] is implemented to estimate the local illuminant color in the first part of the thesis. The median or average of these local estimates represents as a global illuminant color. We propose to incorporate the constraint of neutral region determined by the Planckian locus into the framework of the CNN to boost the performance of color constancy. In the second part of the thesis, the original CNN is retrained and transformed into a classifier which is able to distinguish between high and low color temperature images. The local estimates of the illuminant color from each classified image are fed into the corresponding support vector regressor (SVR) to determine the global illuminant color. In the last part of the thesis, the integration of the CNN classifier and the statistical methods is devised to employ the merits of each statistical method. Experimental results have shown that utilizing the neutral region can greatly elevate the performance of the proposed method. Meanwhile, the results also indicated that the higher the accuracy of the classifier, the lower the estimation error of the illuminant color. HSIN,CHENG-HO 辛正和 2018 學位論文 ; thesis 78 zh-TW |
collection |
NDLTD |
language |
zh-TW |
format |
Others
|
sources |
NDLTD |
description |
碩士 === 逢甲大學 === 通訊工程學系 === 106 === White balance in imaging devices mimics the ability of color constancy in human vision. It assures an accurate color reproduction in images under a variety of illuminant conditions. The first step of white balance is to estimate the color of the illuminant. Then, the estimated illuminant color is applied to eliminate the color cast so that the corrected image appears as taken under the canonical illuminant. The learning based approach in color constancy shows much higher performance than that of the traditional statistical approach. Especially, a convolutional neural network (CNN) enables to take feature extraction into the training process, so it can learn a complex mapping between the input and the output. The purpose of this thesis is to develop methods of estimating illuminant color using CNNs.
The thesis is composed of three parts. The CNN used as a regressor [1] is implemented to estimate the local illuminant color in the first part of the thesis. The median or average of these local estimates represents as a global illuminant color. We propose to incorporate the constraint of neutral region determined by the Planckian locus into the framework of the CNN to boost the performance of color constancy. In the second part of the thesis, the original CNN is retrained and transformed into a classifier which is able to distinguish between high and low color temperature images. The local estimates of the illuminant color from each classified image are fed into the corresponding support vector regressor (SVR) to determine the global illuminant color. In the last part of the thesis, the integration of the CNN classifier and the statistical methods is devised to employ the merits of each statistical method. Experimental results have shown that utilizing the neutral region can greatly elevate the performance of the proposed method. Meanwhile, the results also indicated that the higher the accuracy of the classifier, the lower the estimation error of the illuminant color.
|
author2 |
HSIN,CHENG-HO |
author_facet |
HSIN,CHENG-HO WEI,YIN-JHIH 魏吟芝 |
author |
WEI,YIN-JHIH 魏吟芝 |
spellingShingle |
WEI,YIN-JHIH 魏吟芝 Illuminant Estimation based on Shallow Convolutional Neural Networks |
author_sort |
WEI,YIN-JHIH |
title |
Illuminant Estimation based on Shallow Convolutional Neural Networks |
title_short |
Illuminant Estimation based on Shallow Convolutional Neural Networks |
title_full |
Illuminant Estimation based on Shallow Convolutional Neural Networks |
title_fullStr |
Illuminant Estimation based on Shallow Convolutional Neural Networks |
title_full_unstemmed |
Illuminant Estimation based on Shallow Convolutional Neural Networks |
title_sort |
illuminant estimation based on shallow convolutional neural networks |
publishDate |
2018 |
url |
http://ndltd.ncl.edu.tw/handle/559apg |
work_keys_str_mv |
AT weiyinjhih illuminantestimationbasedonshallowconvolutionalneuralnetworks AT wèiyínzhī illuminantestimationbasedonshallowconvolutionalneuralnetworks AT weiyinjhih jīyúqiǎncéngjuǎnjīshénjīngwǎnglùzhīguāngyuányánsègūcèfāngfǎ AT wèiyínzhī jīyúqiǎncéngjuǎnjīshénjīngwǎnglùzhīguāngyuányánsègūcèfāngfǎ |
_version_ |
1719160015632203776 |