Study On Color Spectral Characteristics Based On Manifold Learning

Our perception of all the things on earth is based on light and color, and the spectral reflectance curve, known as "fingerprint" of a color, is the most accurate way to represent a color. But the dimension of the spectral data is too high to process. Therefore, the problems of number of basic factors and the extracted structure of color space have been paid wide attention.Traditional linear techniques of dimension reduction were applied to these problems, which might lose some important nonlinear information. Besides, the former dimension estimation algorithms tend to overestimate the true dimension. Aiming at these problems, this paper includes research as below:1. This paper provides a new analyzing method different from traditional linear techniques of dimension reduction, which is based on manifold learning, supposing high dimensional data of spectral color lies in a low dimensional manifold, transforming the problems of number of basic factors and the extracted structure of color space to the problems of the intrinsic dimension and structure of the embedded manifold of the spectral color space.2. Five different techniques of intrinsic dimension estimation and six classic manifold learning algorithms are employed to study the spectral dataset of Munsell colors. Experimental results reveal that there exists a3-dimensional manifold embedded in the spectral Munsell color space and the geometric structure of this manifold looks like a cone, consistent with the original development of the Munsell color system.3. Exacting the samples which have the Value of7from Munsell colors, and also experiment with the relationship between dimension one and ME, revealing that dimension one is closely related to Munsell Value, dimension two and dimension three together reflect Munsell Chroma and Hue.4. Based on the Munsell spectral colors experiments, mixing the forest spectra and the spectra of natural and artificial objects to analyze. Different from the former researches, which concluded that more than three basis factors are required in general, the intrinsic dimension of spectral colors should be3as well.