Abstract
The goal of this study is to identify the most fundamental independent elements common to natural color images that can be applied for general image representation, and examine their possible neurophysiological correlates. To this end we applied Independent Component Analysis (ICA) to naturally occurring color images. ICA derives statistically independent image elements by minimizing mutual information. We used images consisting of flowers, leaves, trees, rocks, and other natural objects and background. ICA results in luminance and color filters with simple and complex cell receptive field profiles. The luminance filters are localized and oriented edge detectors as reported previously (1). The color filters comprise of blue-yellow and red-green double-opponent receptive fields with various orientations. ICA shows that the independent components of natural color images are multispectral edges, and predicts that spatio-chromatic information is coded in statistically independent luminance, blue-yellow, and red-green opponent pathways. The color and luminance receptive fields may be considered optimal for edge and orientation detection in color or multispectral images (2). Despite the lack of strong physiological evidence for double-opponent cells in mammalian cortex, ICA supports that decomposition of spatio-chromatic information into luminance, red-green, and blue-yellow channels offers an optimum means of coding natural images. A possibility is that the local computation of double opponency is not carried out by individual cells but distributed across a population of cells.
(1)
BellA. J.SejnowskiT. J.Vision Research, 37, 3327–3338, 1997
(2)
TailorD.FinkelL. H.BuchsbaumG.Vision Research, 40, 2671–2676, 2000