基于BP神经网络的多光谱空间表示方法研究(英文)

Representing Spectral Color Space Using BP Neutral Network

多光谱成像技术通过增加颜色通道的维数,克服了传统颜色复制方式在同色异谱方面的缺陷,成功的实现了基于光谱的颜色复制。然而,由于其颜色信息维数较高,此方法在提高色度精度的同时引入了较大的计算及存储压力。为此,最常用的方法就是通过特定的光谱分组度量对光谱数据进行分组,并利用每组光谱数据集中的主成分向量来对各个光谱曲线进行线性表示,从而实现数据的降维处理。本研究提出了一种全新的,基于主波长分组及BP神经网络寻址的光谱空间表示方法,并通过对具体光谱颜色数据集的向量表示,证明此方法在光谱颜色表达的方面的准确性。

By extending the number of signal channels to high dimensions, the multispectral imaging technology overcomes the shortage of metamerism of traditional color reproduction method and fulfils the spectral color repro- duction. However, it also brings out calculating and storing problem because of the large spectral dimension. The solution for this matter is to divide the database into several subgroups according to certain metrics and find their basic components for each individual subgroup for linear representation. In this research, a new multispectral color space representation workflow was proposed, together with a color group classification method which based on the dominant wavelength grouping and the BP neutral network. This workflow was applied to multispectral color representation in experiment by representing several spectral color datasets, and achieved excellent spectral representation accuracy.