Higher Order Primitives for the Reconstruction of Coarsely Sampled Imagery

Abstract

While computer-processing power continues to grow at rates predicted by Moore's Law, display technologies have lingered and grown at a much slower pace. The ability to display interactive printer-resolution images will not be achievable for several years at the current growth rate of display technologies. To compensate, a method for generating succinct descriptions of hyper-resolution images should be developed. We present such a method that combines sparse color samples with edge information to reconstruct higher resolution images.

Our framework consists of four major phases. In the first phase, we sample the image plane using a sparsely populated grid. We then use a priori knowledge of the scene geometry to detect and record edgelets, each of which is a point sample in the image plane containing the location of a detected geometric edge as well as the edge's orientation. The second phase groups edgelets into continuous contours using a function derived from research on the human visual system and the phenomena of illusory contours. In the third phase, the continuous contours are propagated throughout the higher-resolution reconstruction grid to formulate hypotheses about local image edge structures. In the final phase, each reconstruction point gathers color samples from the coarse grid while respecting hypothesized edge boundaries and subjects the samples to a Gaussian filter to determine its contribution to the reconstructed color.

Although more refinements are necessary for ideal reconstructions, results indicate that combining the sparse edge information with sparse color samples is expressive enough to reconstruct images with crisp edge boundaries, even under conditions of extreme under sampling.