Learning contour statistics from natural images

2012-11-13Vision is one of our most important senses, and accordingly, a large fraction of our cerebral cortex is devoted to visual processing.One of the key computations in the early stages of the visual system is the extraction of object contours, since the occluding boundaries and orientation discontinuities of objects that make up the ""line drawing"" of a scene are the most important and direct cues to object shape. Contour detection in natural scenes has proven to be a difficult technical problem, however, mainly because the existence (or not) of a contour at a given location, orientation and scale depends, probabilistically, on a large number of cues covering a large fraction of the visual field.Worse, the cues interact with each other in a multitude of ways and combinations, leading to an enormously complex cue integration problem. ❧ In this work, we attempt to break down the contour extraction problem into natural, tractable, modular subcomputations.In chapter 2, we describe a novel approach to combining local edge cues from the area generally orthogonal to the orientation of the contour.Key aspects of the approach are the (1) tabulation and modeling of contour statistics at fixed values of local edge contrast, to reduce higher-order dependencies within the population of local edge cues, and (2) picking the most informative contour cues from the de-correlated local edge population. The resulting contour operator has no parameters and has significantly improved localization and sharpened orientation tuning compared to the raw local filter values. In chapter 3, we describe a novel approach to combining cues from the area generally tangential to the contour.In this case, we have developed an approach to efficiently gather the contour statistics needed to optimally use ""contextual"" cues from aligned high-resolution flankers and the superimposed coarse-scale edges.In the process, we have found evidence that the integration of these contextual cues across scales can be achieved by a cascade of simple 2-input functions, greatly simplifying our statistics-driven approach.We found that the interaction between two collinear flankers is similar to a minimum like operation, center response and flankers interaction is conjunctive/contextual, and for the particular case of cross scale interaction we looked at there was minimal interaction.We generalized the approach developed for collinear cues to cues for curved contours.The resulting contextually-boosted contour operator strongly emphasizes spatially-extended contours found in natural scenes, again with zero parameters.We also describe a novel image enhancement method based on Cornsweet illusion using contours obtained from the above two methods