Edge-Preserving Laplacian Pyramid

5th International Symposium on Visual Computing, Las Vegas, Nevada, USA, Nov 30 Dec 2, 2009 The Laplacian pyramid recursively splits an image into local averages and local differences using a fixed Gaussian interpolation function. We propose a spatially variant interpolation function that is adaptive to curvilinear edges in the image. Unlike the signal-based multiscale analysis where a step edge is multiply represented at all scales, our perception-based multiscale analysis preserves the edge at a single scale as much as possible. We demonstrate that our average pyramid retains boundaries and shading at lower spatial and tonal resolutions, whereas our difference pyramid refines edge locations and intensity details with a remarkably sparse code, delivering an image synopsis that is uncompromising between faithfulness and effectiveness.