Document image dewarping using Kinect depth sensor

Abstract

This paper presents a novel approach for document de-warping using Microsoft Kinect camera sensor. The proposed method makes use of the 3D model of the warped document using Kinect sensor. Due to the low resolution of Kinect's RGB camera, an external high resolution camera is used for capturing a higher quality image. The point cloud captured by Kinect depth sensor is down sampled and then the voxilized grid approach is applied. Then a triangulation mesh is generated using greedy fast triangulation method. In the next step, a conformal 3D to 2D mapping is applied to the curved mesh. This creates a flat plane of the 3D mesh. Eventually, the original image is converted to the flatted version by a 2D to 2D mapping method. Text independency of this method distinguishes it from the other methods which use the document text lines to estimate the curvature of the plane. By using this method, documents with different distortions and with any kind of content such as pictures, graphs and tables can be de-warped.