Efficient Retina-like Resampling from Cartesian Images

This paper describes a novel method to resample cartesian images to a retina-like hexagonal tessellation, consisting of foveal and peripheral regions. We introduce the inflated-hexagons model for the foveal region resampling, which has an efficient data structure for storage and mapping of cortical coordinates. The integral image representation is used as intermediate step for the peripheral region log-polar resampling, which is less computationally demanding than the simulation of Gaussian receptive fields with varying widths. An important feature of the resulting model is that a gap-free transition is obtained between the near-uniformly sampled fovea and the space-variant periphery. Experiments with Principal Component Analysis (PCA) compression of the ORL Database of Faces show that the proposed retina-like resampling, in spite of massively reducing the amount of image data, still retains most of the information contents of conventional cartesian image sampling.