In recent years, many documents in the libraries are electronically prepared and can be stored directly in a digital archive for reference, search, and distribution. Particularly, for the collections of historically significant and in some cases badly deteriorated documents, the primary purpose of digitization is the image itself, which serves as a photo-realistic facsimile. Google, which has the world’s largest search engine database, recently announced its plan to digitize historical documents and out-of-print books in selected libraries. A common problem that comes with document digitization is the document distortion. Deformations due to handling or mishandling documents cause the damage and bucking of the documents. Years of folding of documents can also leave them rigidly warped and creased. Other causes include the natural bending of pages, especially near a book’s binding. The conventional solution to the restoration question is direct repair of the actual physical artifact. However, such physical restoration suffers from the following drawback: The materials to be restored may be rare and singularly unique. Their conditions can be so fragile that physical restoration, especially procedures that must alter shape, can pose a great risk for further and possibly irreversible damage. Given this risk and the subjective nature of the restoration process itself, offering a digitally-based solution becomes very attractive.In digital restoration of distorted documents, the previous methods limit themselves in the area of 2D image enhancement. These techniques can be used to help increase the perceptual saliency of features in digitized material, such as characters in a text. However, the pure 2D image operations can not address shape-based distortions without additional constraints. The pinhole camera for document imaging is a projective transformation engine, and by its nature creates an image with systematic projective distortion. The unique imaging orientation that does not produce projective distortion occurs when all points on the objects have the same distance from the camera, as is the case for truly flat manuscripts imaged in an orthogonal orientation to the camera’s optical axis. The 3D shape distortion on the surface of the manuscript and the imaging process which produces projective distortion are confounded in the image to yield a complex 2D distortion of the text of a manuscript that was intended originally to be flat. The previous 2D image enhancement methods ignore projective distortion. As a result, the projective distortion remains in an image after 2D restoration and enhancement. Because it is difficult to estimate surface shape from a single image, there is no reliable way to identify and remove projective distortion via 2D image processing.
The objective of our restoration is to create a planar representation of an originally planar document that has undergone an arbitrary and unknown deformation. We developed a system that can acquire and flatten the 3D shape of a warped document to determine a nonlinear image transformation that can correct for image distortion caused by the document’s shape. Our method is to use a 2D image of a document together with its 3D shape in order to restore it by removing the projective distortion present in the image. Our algorithm employs a dynamic deformable model based on the MSD system for physically-based modeling of acquired 3D surface of the document. After the distorted surface of the document is modeled by the MSD system, it can be dynamically deformed towards a plane under a field of external and external force vectors. The “flattening?deformation is achieved by calculating the energy equilibrium state of the MSD mesh based on the Lagrange’s dynamics.
Snapshots of the simulation of flattening the textured mesh.
Left column: original distorted 2D images. Central column: extracted shading image before geometric
restoration. Right column: final restored image with both geometric and photometric distortions removed.
(a) A multi-folded page; (b) Original 2D warped image; (c) Sub-sampled sparse triangular mesh (2,909 points); (d)
Sparse quad mesh with shear springs (999 points); (e) Restored quad mesh with a previous mass-spring model; (f) Restored
2D image of (e); (g) Restored result using our triangular mesh with stick model; (h) Extracted intrinsic shading image; (i)
Reflectance image with shadings removed.
(a) A crumpled document image; (b) Original triangular mesh with 5,000 points (bending resistance shown); (c)
Geometrically restored image; (d) Extracted inpainting mask; (e) Extracted shading image; (f) Final reflectance image with
shadings removed.
K.B. Chua, L. Zhang, Y. Zhang and C.L. Tan. "A fast and stable approach to restore warped document images". Eighth International Conference on Document Analysis and Recognition (ICDAR 2005), pp. 384-388, Seoul, Korea, 29 Aug - 1 Sept
2005.
Copyright 2005-2013, Yu
Zhang. This material may not be published, modified or otherwise
redistributed in whole or part without prior approval.
