基于主成分分析的彩色扫描仪光谱特性化

为了实现扫描仪在不同光源、不同观察者条件下准确获取颜色信息,最大程度的避免同色异谱现象,本文采用光谱的方法对扫描仪进行特性化处理,通过多项式回归和BP神经网络分别与主成分分析法结合,首先对检测样本的光谱反射率进行主成分分析,提取主成分与主成分系数,通过实验得到主成分系数与多项式回归、BP神经网络结构之间的转换模型,实现了扫描仪低维RGB信号对原始光谱反射率信息的重构,进而实现扫描仪的光谱特性化.实验结果表明,多项式项数为19项时,达到训练样本的均方根误差为1.7%,检测样本的均方根误差为1.9%.而包含15个隐层节点的单隐层BP神经网络结构为比较合理的网络结构,达到训练样本的均方根误差为1.3%,检测样本的均方根误差为1.5%.对彩色扫描仪的特征化处理,采用多项式回归法得到光谱特性化精度较低,采用BP神经网络模型能够实现更高的光谱特性化精度.

Spectral Characterization of Color Scanners Based on Principal Component Analysis

Spectral characterization of color scanners can achieve accurate obtaining of color information of scanners in the different light sources and observers. It can also avoid metamerism to the most extent. Spectral characterization was applied to characterize the color scanners. Principal component analysis, which was combined with polynomial regression and BP neural network technology, set up the nonlinear transformation relationship between the scanner RGB signal and spectral reflectance image information. Firstly, principal component analysis (PCA) was used to analyze the spectral reflectance of the training sample followed by the principal component scalars calculation and the spectral was represented by principal component times scalars of principal component. Conversion models for scalars of principal component and polynomial or BP neural network was built by experiments. The reflectance was built by RGB low-dimension signal and spectral characterization of color scanners was achieved. Experimental results showed that the better number of polynomial terms was 19, reaching the accuracy of 1.7% RMSE of training sample and 1.9% RMSE of test sample. And the optimum network structure was single hidden layer with 15 layer node, reaching the accuracy of 1.3% RMSE of training sample and 1.5% RMSE of test sample. The accuracy of polynomial regression method is much lower, which is not fit for spectral characterization. The BP neural network model may achieve higher spectral characterization accuracy.