Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/1281
Title: | 基於卷積神經網路與色彩感知的水彩混色模型 Perceptual-Based CNN Model for Watercolor Mixing Prediction |
Authors: | Ya-Bo Huang 黃雅博 |
Advisor: | 歐陽明(Ming Ouhyoung) |
Keyword: | 顏色匹配函式,顏料混色,光譜,卷積神經網路,機器學習, Color Matching Function,Pigment Mixing,Spectrum,Convolutional Neuron Network,Machine Learning, |
Publication Year : | 2018 |
Degree: | 碩士 |
Abstract: | 本論文主要探討將卷積神經網路(CNN)應用於水彩混色預測的模型,以及利用色彩感知誤差作為損失函式對於模型訓練的影響。目前顏料混色之研究多以Kubelka-Munk理論作為基礎,然而K-M理論本身的限制,無法完全符合半透明的水彩顏料預測之需求。因此,設計一個更符合的水彩顏料的混色模型,能夠對未來水彩模擬以及相關應用有所貢獻。本文用一多層卷積神經網路來學習兩顏料混色後的反射頻譜。訓練資料部分使用[Chen et al. 2018]之水彩資料庫,訓練目標是最小化混色後的反射頻譜在上與真實混色之色彩誤差。透過本文所提出的模型以及損失函式,混色預測在測試資料中,有88.7%的結果能夠達到∆ELab < 5的色彩誤差,也就是人眼無法輕易分辨其色差之程度。 此外,本研究發現一發生在預測結果上之特別現象。此模型透過色感誤差的損失函式訓練後所產生的結果在頻譜上,相較於真實混色,會有中幅度的波動,無法貼近真實頻譜。然而在色彩上,卻因為損失函式,而能在色彩上與真實混色貼近。 In the paper, we propose a model to predict the mixture of watercolor pigments using convolutional neural networks (CNN). With a watercolor dataset, we train our model to minimize the loss function of sRGB differences. In metric of color difference ∆ELab, our model achieves 88.7% of data that ∆ELab < 5 on the test set, which means the difference cannot easily be detected by the human eye. In addition, an interesting phenomenon is found; Even if the reflectance curve of the predicted color is not as smooth as the ground truth curve, the RGB color is still close to the ground truth. |
URI: | http://tdr.lib.ntu.edu.tw/handle/123456789/1281 |
DOI: | 10.6342/NTU201804120 |
Fulltext Rights: | 同意授權(全球公開) |
Appears in Collections: | 資訊工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-107-1.pdf | 1.63 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.