Local image descriptors provide robust descriptions of image localities. Their geometric arrangement provides additional information about the image they describe, a fact often ignored when employing them to that wide slew of tasks from image registration to scene classification. On the premise that descriptor quality could be assessed in terms of its expressiveness of image content, we investigate the use of the described as well as that additional geometric information to the task of recovering the image from its local descriptors. This paper uses Local Binary Patterns, an operator nested in a dense geometry, to study how this additional information in the form of constraints among pixels dictates the intensity estimated for a pixel. We determine that constraints propagate from regional extrema to regions around them that observe the same constraint class, and that the intensity for any of the region's pixels influences that for all others. We build a directed constraint graph of pixel nodes such that the arcs on the graph are strongly k-consistent, and propagate intensity estimates from extremum nodes. Evaluations are run on the SIPI texture and the BSD500 datasets. The estimates preserve the local structure of the image, as shown by the Mean Absolute Error of about 15% and 18% respectively and Structural Texture SIMilarity of about 92% for both datasets, in addition to observing 100% constraint satisfaction.
展开▼