An L_1 Image Transform for Edge-Preserving Smoothing and Scene-Level Intrinsic Decomposition

摘要

Identifying sparse salient structures from dense pixels is a longstandingrnproblem in visual computing. Solutions to this problemrncan benefit both image manipulation and understanding. In this paper,rnwe introduce an image transform based on the L_1 norm forrnpiecewise image flattening. This transform can effectively preservernand sharpen salient edges and contours while eliminating insignificantrndetails, producing a nearly piecewise constant image withrnsparse structures. A variant of this image transform can performrnedge-preserving smoothing more effectively than existing state-of the-rnart algorithms. We further present a new method for complexrnscene-level intrinsic image decomposition. Our method relies onrnthe above image transform to suppress surface shading variations,rnand perform probabilistic reflectance clustering on the flattened imagerninstead of the original input image to achieve higher accuracy.rnExtensive testing on the Intrinsic-Images-in-the-Wild database indicatesrnour method can perform significantly better than existingrntechniques both visually and numerically. The obtained intrinsicrnimages have been successfully used in two applications, surface retexturingrnand 3D object compositing in photographs.