Noam Elron (The Hebrew University of Jerusalem)
Tuesday, 27.1.2015, 11:30
The local self-similarity (LSS) image prior provides a reliable and efficient source for example patches in natural images. We present a new image upscaling method in which the LSS prior is: (i) fulfilled exactly along image singularities, e.g., edges, and (ii) its use is maximized where it holds and suppressed otherwise. The new approach interleaves the pixels of several pre-scaled images, which differ by sub-pixel offsets, to form the low-frequency layer of the output image. The interleaving process ensures that every pixel or patch in the output grid has a geometrically-aligned counterpart in the low-resolution example images. Thus, we use the LSS in a strict sense and fill in the missing output high-frequency layer by fusing together predetermined example patches. Thus, we avoid the need to perform patch search and post-correction operations which current upscaling methods do. The new scheme thus offers greater efficiency and reconstructs edges more faithfully in terms of their sharpness and without introducing halos or staircasing artifacts.
We further refine the LSS image prior and show that it is concentrated along narrow veins of image singularities. We propose a simple measure for the validity of the prior and use it for two purposes. Instead of averaging together overlapping example patches, we prefer patches with higher LSS score. At regions where the score is low, i.e., the LSS prior does not hold sufficiently, we suppress its use. The new measure is light to compute, and allows us to both achieve sharper edges as well as reduce the facet-like artifacts that result from improper use of this prior at textured regions.