Segmentation and Rectification of Pictures in the Camera-Captured Images of Printed Documents

This paper presents an algorithm that segments and rectifies pictures in camera-captured document images. Most of the conventional methods for this purpose require the 3-D shape of document surface, which are usually measured or inferred by a depth-measuring device, structured light, or stereo system. Unlike these methods, our method requires only a single-view image and a user-provided rough bounding box on the picture. Hence, the main features of the proposed algorithm are simple user interaction and short processing time: a mega-pixel size image can be segmented and rectified within 1-2 s, on receiving the user's bounding box. To achieve this goal, we develop a novel boundary extraction algorithm that exploits the specific properties of printed material. In the method, a set of boundary candidates is generated, and the optimal boundary is found by using an alternating optimization scheme. In addition to the segmentation method, we also propose a new rectification method, which can largely remove perspective distortions. Experimental results on a variety of images show that our method is efficient, robust, and easy to use.