Fast randomized multi-scale energy minimization for image processing

摘要

An image processing module performs efficient image enhancement according to a multi-scale energy minimization process. One or more input images are progressively downsampled to generate a pyramid of downsampled images of varying resolution. Starting with the coarsest downsampled image, a label map is generated that maps output pixel positions to pixel positions in the downsampled input images. The label map is then progressively upsampled. At each upsampling stage, the labels are refined according to an energy function configured to produce the desired enhancements. Using the multi-scale energy minimization, the image processing module enhances image via hole-filling and/or super-resolution.