Doctoral Dissertation Video Mosaicing Based on Structure from Motion for Geometric Distortion-Free Document Digitization

This thesis describes a novel video mosaicing method which is capable of generating a geometric distortion-free mosaic image. Video mosaicing, which stitches partial images of a target into a larger image called a mosaic image, has been used to obtain high resolution images of documents. Generally, mosaic images are prone to two types of geometric distortions: perspective distortion which appears when the target document is not fronto-parallel to the camera’s image plane, and curvature distortion which is caused by projecting curved surface of the target document to the image plane of the camera. This thesis first focuses on a flat document, and proposes a perspective distortion-free video mosaicing method for flat documents. In the proposed method, extrinsic camera parameters are estimated for each frame by applying structure from motion technique to the captured video. Using estimated extrinsic camera parameters, the method dewarps all the frame images and synthesizes them on a virtual fronto-parallel plane to generate a super-resolved mosaic image without perspective distortion. Then, this method for flat documents is extended to deal with curved documents. This extended method generates a virtually flattened mosaic image of a curved surface. In this method, first, extrinsic camera parameters, along with ∗Doctoral Dissertation, Department of Information Systems, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0661201, August 21, 2008.