Removing shadows from images of documents

Abstract

In this work, we automatically detect and remove distracting shadows from photographs of documents and other text-based items. Documents typically have a constant colored background; based on this observation, we propose a technique to estimate background and text color in local image blocks. We match these local background color estimates to a global reference to generate a shadow map. Correcting the image with this shadow map produces the final unshadowed output. We demonstrate that our algorithm is robust and produces high-quality results, qualitatively and quantitatively, in both controlled and real-world settings containing large regions of significant shadow.