Image analysis through local information measures

Abstract

The properties of local image statistics are analyzed in a classic information theoretic setting. Local spatiochromatic image elements are projected into a space in which constituent components are independent by way of independent component analysis, allowing a fast and tractable means of considering the joint likelihood of such statistics. Observation of this likelihood allows inferences to be made regarding the informativeness of a particular set of statistics. This operation is shown to illuminate a number of perceptually important image properties, allowing figure-ground segmentation, removal of common or expected image elements, and prediction of regions of interest.