The Spectral Independent Components of Natural Scenes

We apply independent component analysis (ICA) for learning an efficient color image representation of natural scenes. In the spectra of single pixels, the algorithm was able to find basis functions that had a broadband spectrum similar to natural daylight, as well as basis functions that coincided with the human cone sensitivity response functions. When applied to small image patches, the algorithm found homogeneous basis functions, achromatic basis functions, and basis functions with overall chromatic variation along lines in color space. Our findings suggest that ICAm ay be used to reveal the structure of color information in natural images.