Quadrilateral Signboard Detection and Text Extraction.

Abstract

There are numerous rectangular signboards indoors and outdoors. The information contained in them, such as signs and text, is useful for robot navigation. In this paper, a system for automatically detecting signboards and extracting text from them is presented. First it locates the signboard’s position on the image by the Hough transform, edge density checking and quadrilateral finding. Then, with a geometric transformation, the system changes the signboard from an arbitrary quadrilateral to a rectangular shape. Finally, it performs segmentation to extract the characters on the signboard. In our system, the signboards can have large variations such as rotation and scaling and are not limited to be (nearly) rectangular in the images. Experiments show that our system is quite effective.