State-of-the-art scene text detection methods usually model the text instance with local pixels or components from the bottom-up perspective and, therefore, are sensitive to noises and dependent on ...