摘要
颜在日常生活中随处可见，与我们生活息息相关。随着科学技术的不断发展，人们对彩图像的颜复制的要求也随着提高，颜复制技术现已成为颜工作者研究的热点。目前，颜复制技术主要有两种，貌模型的颜复制和光谱的颜复制。前者复制后的彩在视觉上具有相同的颜，光谱可能并不一致，无法避免由同异谱现象带来的困扰。光谱颜复制是通过使用光谱反射率作为颜信息传递和再现的媒介，能确保颜的一致性，现已被广泛使用。传统获取光谱反射率的方法主要有分光光度计和高光谱分析仪，但由于两者自身存在的局限性，常在实际应用中造成不必要的麻烦。为此，颜科技工作者们提出用多光谱成像技术来获取光谱反射率，利用多光谱成像系统相机输出的多通道图像信息对目标样本的光谱反射率进行估算，估算的过程被称之光谱重建。彩数码相机能够在多种条件下，对目标样本采取非接触式成像，且数码相机具有灵活方便性、性价比高等优势，因此本文采取彩数码相机搭载一套滤光片组成多光谱成像系，统获RGB信号重建出物体表面反射率。
当前光谱重建得到物体表面光谱反射率估值的常用方法有伪逆法、主成分分析法和BP神经网络算法。基于BP神经网络重建算法的不足，引入多项式模型和贝叶斯正则化修正项，改进传统的BP神经网络光谱重建算法，以此来优化算法提高精度。为了便于验证重建算法的可行性，实验中训练样本使用标准卡Digital ColorChecker SG，检验样本使用标准卡Color Checker Rendition Chart。实验结果表明本文所提出的算法重建的光谱反射率，无论在度精度还是光谱精度上都优于传统神经网络算法，且精度远远大于伪逆
法，主成分分析法，说明本文所提出的方法对物体表面颜的真实再现具有一定的价值。
最后利用本文提出的贝叶斯正则化神经网络光谱重建算法，得到重建后的光谱反射，再结合度学和计算机图形图像学知识重现蜡染画芯真实彩。对画芯中的同一块，在光源不同角度下，对其重现后的颜块采用HSI和L*a*b*颜空间分析。实验结果表明，HSI颜空间的稳定性更好，能为后续参考标准块的任一实拍图像拟合出任意角度下的图像工作，提供理论支撑。
关键词：多光谱成像；光谱重建；神经网络算法；贝叶斯正则化；多角度；颜空间
I
Abstract
Color is everywhere in daily life and is closely related to our lives. With the continuous development of science and technology, people's requirements for color reproduction of color images have also increased. Color reproduction technology has become a hotspot for color workers. At present, there are mainly two types of color reproduction technology, color reproduction of color appearance models and color reproduction of spectra. The former copied color has the same color visually, and the spectrum may not be consistent, which can not avoid the trouble caused by metamerism. Spectral colo
r reproduction is the use of spectral reflectance as a medium for color information transmission and reproduction to ensure color consistency, and is now widely used. Traditional methods for obtaining spectral reflectance mainly include spectrophotometers and hyperspectral analyzers, but due to the limitations of the two, they often cause unnecessary trouble in practical applications. For this reason, color science and technology workers have proposed to use multispectral imaging technology to obtain spectral reflectance, and use the multi-channel image information output by the multispectral imaging system camera to estimate the spectral reflectance of the target sample. The process of estimation is called spectral reconstruction. Color digital cameras can take non-contact imaging of target samples under a variety of conditions, and digital cameras have the advantages of flexibility, convenience, and high cost performance. Therefore, this paper adopts a color digital camera equipped with a set of filters to form a multispectral imaging system. The signal reconstructs the surface reflectance of the object.
At present, the commonly used methods for estimating the spectral reflectance of an object's surface by spectral reconstruction include pseudo-inverse method, principal component analysis method and BP neural network algorithm. Based on the shortcomings of the BP neural network reconstruction algorithm, a polynomial model and a Bayesian regularization correction term are introduced to improve
the traditional BP neural network spectral reconstruction algorithm to optimize the algorithm and improve its accuracy. In order to verify the feasibility of the reconstruction algorithm,
II
the training samples used the standard color card Digital Color Checker SG, and the test samples used the standard color card Color Checker Rendition Chart. The experimental results show that the spectral reflectance reconstructed by the algorithm proposed in this paper is superior to traditional neural network algorithms in terms of chrominance accuracy and spectral accuracy, and the accuracy is far greater than that of the pseudo-inverse method and principal component analysis method. The method has certain value for the true reproduction of the surface color of the object.
Finally, the Bayesian regularized neural network spectral reconstruction algorithm proposed in this paper is used to obtain the reconstructed spectral reflection, which is combined with colorimetry and computer graphics imaging knowledge to reproduce the true reproduction of the color of the batik core. For the same color block in the picture core, under different angles of the light source, the color block after its reproduction is analyzed by HSI and L*a*b* color space. The experimental results show that the HSI color space has better stability, and can provide theoretical support for fitting any real-time image of the reference standard color block to any image work at any angle.
Key Words: Multispectral imaging; spectral reconstruction; neural network algorithm; Bayesian regularization; multi-angle; color space
III
目录
摘要 ........................................................................................................ I I 目录 ......................................................................................................... IV 第一章绪论 (1)
1.1 论文研究背景及研究目的 (1)
1.2 国内外研究现状概述 (2)
1.2.1 多光谱成像技术的发展 (2)
1.2.2 光谱反射率重建算法的研究进展 (6)
1.3 论文研究的主要内容 (8)
第二章颜科学基本理论 (9)
2.1 颜的基本原理 (9)
2.1.1 颜的感知 (9)
2.1.2 标准光源 (10)
2.1.3 物体颜的光谱特性 (11)
2.2 CIE标准度系统 (12)
2.2.1 CIE 1931 RGB度系统 (12)
2.2.2 CIE 1931 XYZ度系统 (14)
2.2.3 CIE1976 L*a*b*颜空间 (16)
2.2.4 CIEDE2000差 (18)
2.3 sRGB彩空间 (20)
2.4 同异谱 (20)
2.5 颜再现原理 (21)
2.6 本章小节 (22)
第三章光谱反射率重建算法 (23)
IV
3.1 多光谱成像系统 (23)
3.2 光谱重建算法 (25)
3.2.1 伪逆法 (26)
3.2.2 主成分分析法 (27)
3.2.3 BP神经网络算法 (28)
3.3 本章小节 (31)
第四章基于贝叶斯正则化神经网络的光谱重建 (32)
4.1 贝叶斯正则化光谱重建算法 (32)
4.1.1 多项式回归法 (32)
4.1.2 贝叶斯正则化神经网络 (33)
4.2 实验仪器及实验过程 (34)
4.2.1 实验器材 (34)
4.2.2 数据采集 (36)
4.2.3 实验数据的计算 (38)
4.2.4 光谱反射率重建精度的评价方法 (39)
4.3 结果分析 (40)
4.4 本章小节 (47)
第五章不同角度下多光谱成像系统光谱反射率重建研究 (48)
5.1 颜空间 (48)
5.1.1 HSI颜空间 (49)
5.1.2 CIE1976 L*a*b*颜空间 (50)
5.2 仪器设备及实验样本的获取 (50)
5.2.1 光源的选择 (50)
5.2.2 实验样本和设备 (51)
5.3 实验结果及其分析 (55)
5.3.1 重建各个照明角度下各块图像 (55)
5.3.2 块在HSI颜空间中随光源角度的变化情况 (56)
5.3.2 块在CIE1976 L*a*b*颜空间中随光源角度的变化情况 (59)
正则化网络
V

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。