Geometry-Guided Non-Local Transform-Domain Filtering and Coherent Optical Depth-of-Field Synthesis via Joint Monocular Disparity and Surface Normal Fields

Defensive publication (technical disclosure) placing the disclosed subject matter into the public domain as prior art. Abstract:A single photograph, possibly captured at high sensor gain, is processed by a neural geometry estimator that predicts, from that one image, a pixel-aligned pair of fields: a scalar disparity (or depth) field and a field of unit surface normals in camera coordinates. The pair drives two consumers:(i) A non-local, transform-domain collaborative filter of the block-matching family groups similar patches only when they agree photometrically and geometrically: the patch distance is augmented with a depth term, a normal-orientation term and a boundary-crossing term, applied as a hard rejection before grouping, as a soft weight during aggregation, or both; the search window and the patch support are further deformed along the tangent plane given by the normal field, so that groups never straddle depth discontinuities or geometric creases.(ii) A depth-of-field synthesizer converts the refined disparity into a circle-of-confusion radius with the thin-lens model, separates the image into occlusion-ordered layers, extends each background layer behind its occluders before blurring, and convolves each layer with an aperture kernel of arbitrary shape. Both consumers may share one geometry inference. The document gives the general formulation, the operators, pseudocode and a systematic list of variants. Author / Publisher: INTERVIOLATE (https://interviolate.com, [email protected])Document Version: 1.0.0Document Type: Defensive Publication / Technical DisclosureSuggested CPC Classes: G06T5/70, G06T5/20, G06T5/94, G06T7/50, G06T7/55, G06T2207/20084, G06T2207/20012, H04N25/60, H04N23/80, H04N23/958.