Daylight Spectrum Estimation from Hyper-and Multispectral Image without Area Extraction of Uniform Materials

This paper presents a method for estimating daylight spectrum accurately from a hyperspectral (HS)/multispectral (MS) image without area extraction of uniform materials. To identify materials from spectra obtained with an HS/MS camera in outdoor environments, the daylight spectrum in the scene needs to be estimated and removed. A major conventional method for estimating an illumination spectrum based on the dichromatic reflection model requires extracting multiple areas each covering a highlighted uniform material from images beforehand. The proposed method employs a daylight spectrum model and estimates the daylight spectrum as a point within a model subspace in the spectral space, which minimizes the distance from the hyper-plane on which observed spectra distribute. Experimental results with images of rippled water surface show that the proposed method successfully estimates daylight spectrum without the area extraction. The estimation error of the proposed method evaluated with spectral angle mapper is 2.75° on average, which is almost equivalent to the intrinsic error of the daylight model itself, whereas that of the conventional method is 6.83°.