Text Extraction From Texture Images Using Masked Signal Decomposition

摘要

Text extraction is an important problem in image processing with applicationsfrom optical character recognition to autonomous driving. Most of thetraditional text segmentation algorithms consider separating text from a simplebackground (which usually has a different color from texts). In this work weconsider separating texts from a textured background, that has similar color totexts. We look at this problem from a signal decomposition perspective, andconsider a more realistic scenario where signal components are overlaid on topof each other (instead of adding together). When the signals are overlaid, toseparate signal components, we need to find a binary mask which shows thesupport of each component. Because directly solving the binary mask isintractable, we relax this problem to the approximated continuous problem, andsolve it by alternating optimization method. We show that the proposedalgorithm achieves significantly better results than other recent works onseveral challenging images.