Bilateral Guided Upsampling

摘要

We present an algorithm to accelerate a large class of image processingrnoperators. Given a low-resolution reference input and outputrnpair, we model the operator by fitting local curves that map therninput to the output. We can then produce a full-resolution output byrnevaluating these low-resolution curves on the full-resolution input.rnWe demonstrate that this faithfully models state-of-the-art operatorsrnfor tone mapping, style transfer, and recoloring. The curves arerncomputed by lifting the input into a bilateral grid and then solvingrnfor the 3D array of affine matrices that best maps input color tornoutput color per x, y, intensity bin. We enforce a smoothness termrnon the matrices which prevents false edges and noise amplification.rnWe can either globally optimize this energy, or quickly approximaterna solution by locally fitting matrices and then enforcing smoothnessrnby blurring in grid space. This latter option reduces to joint bilateralrnupsampling [Kopf et al. 2007] or the guided filter [He et al. 2013],rndepending on the choice of parameters. The cost of running thernalgorithm is reduced to the cost of running the original algorithm atrngreatly reduced resolution, as fitting the curves takes about 10 msrnon mobile devices, and 1-2 ms on desktop CPUs, and evaluating therncurves can be done with a simple GPU shader.