Rapid, Detail-Preserving Image Downscaling

摘要

Image downscaling is arguably the most frequently used image processingrntool. We present an algorithm based on convolutional filtersrnwhere input pixels contribute more to the output image the morerntheir color deviates from their local neighborhood, which preservesrnvisually important details. In a user study we verify that users preferrnour results over related work. Our efficient GPU implementationrnworks in real-time when downscaling images from 24M torn70 k pixels. Further, we demonstrate empirically that our methodrncan be successfully applied to videos.