Perceptual color classification based on lightning environment with hyperspectral data

Abstract

Identifying objects by its color name is an essential capability for a service robot to understand and interact with the physical world and be of use in everyday life scenarios. The robot uses an image for the identification, but the image is strongly affected by lightning environment while human does not affected by the environment when they decide its color name. We present lightning environment estimation and perceptual color classification based on the estimation using hyperspectral data. Support Vector Machine (SVM) or Multiple Kernel Learning SVM is used for the estimation and classification. Originality of our paper is that the hyperspectral data is used for both lightning environment estimation and perceptual color classification of an object. Suppose that the lightning environment is given, the perceptual color of the object is classified in each environment. Additionally, perceptual color is estimated by using result of the lightning environment estimation. A novel dataset which consists of the hyperspectral image of 15 objects taken in 4 different lightning environments is constructed in our experiment. Experimental result of the hyperspectral data are compared by those of common color spaces, such as RGB and L*a*b. The estimation and classification with SVM and MKL SVM are compared in the experiments. From the experimental results, hyperspectral data enabled us to present a separated scheme which consists of the lightning environment estimation and perceptual color classification while previous color spaces could not.