A Novel Method of Detecting Chinese Rendered Text on Tilted Screen of Mobile Devices

Abstract

Due to the development of the Internet, screen rendered text detection and recognition have wide applications such as screen capture translation, medical text reading, screen text assisted reading and so on. A large number of screen-rendered attentions have been paid to the English text detection, multi-language scene detection and indefinite length recognition. However, these studies mainly focus on getting the screen rendered image from the device itself. In the actual scene, lots of screen rendered text is mainly captured by the mobile device, and text image has a certain capturing angle. Because of the tilted text line, the current method has a severe degradation in detection performance. This paper proposes a Chinese-style scene text detection method for tilting screens capturing by mobile devices. Firstly, in order to obtain robust detection performance, the Connectionist Text Proposal Network (CTPN) method is employed for horizontal text detection. Secondly, to overcome the dependence of detection on horizontal text, the lightweight Oxford Visual Geometry Group network (VGGnet) is used to correct the tilted direction of the whole image. Moreover, the word width of special characters and punctuation characters are far below average value, so this paper uses the Gaussian removal method to delete the narrow word width and designed a more accurate network to recognize characters. Experiments show that based on the comprehensive scheme, this paper achieves excellent results, even in some extreme bad cases.

Keywords

• Tilted text detection
• mobile application
• character segmentation
• optical character recognition