Wu Weijia ; Xing Jici ; Zhou Hong

Abstract

In this paper, we propose a pixel-wise method named TextCohesion for scenetext detection, which splits a text instance into five key components: a TextSkeleton and four Directional Pixel Regions. These components are easier tohandle than the entire text instance. A confidence scoring mechanism isdesigned to filter characters that are similar to text. Our method canintegrate text contexts intensively when backgrounds are complex. Experimentson two curved challenging benchmarks demonstrate that TextCohesion outperformsstate-of-the-art methods, achieving the F-measure of 84.6% on Total-Text andbfseries86.3% on SCUT-CTW1500.