Scene text removal via cascaded text stroke detection and erasing

Abstract Recent learning-based approaches show promising performance improvement for the scene text removal task but usually leave several remnants of and provide visually unpleasant results. In this work, a novel end-to-end framework is proposed based on accurate stroke detection. Specifically, problem decoupled into detection removal; we design separate networks to solve these two subproblems, latter being generative network. These are combined as processing unit, which cascaded obtain our final model removal. Experimental results demonstrate that method substantially outperforms state-of-the-art locating erasing text. A new large-scale real-world dataset with 12,120 images has been constructed made available facilitate research, current publicly datasets mainly synthetic so cannot properly measure different methods.