Exploiting Spatial and Spectral Image Regularities for Color Constancy

摘要: We study the problem of color constancy–inferring from an image spectrum illumination and reflectance spectra all depicted surfaces. This estimation is underdetermined: many surface can be described as a linear combination 3 basis functions, giving unknowns per pixel, plus parameters for global illumination. A trichromatic visual system makes fewer measurements than there are unknowns. address this by writing small groups pixels combinations ”spatio-spectral” functions. These aggregated require to describe sum spectral individual pixels, us more unknown parameters. explore in Bayesian context, showing when over or underdetermined based on analyzing local curvature characteristics log-likelihood function. show how using spatio-spectral functions might give improved estimates applied real data.