Registration and categorization of camera captured documents
Department of Computer Science
Doctor of Philosophy
Shih, Frank Y.
McHugh, James A.
Camera captured documents
Camera captured document image analysis concerns with processing of documents captured with hand-held sensors, smart phones, or other capturing devices using advanced image processing, computer vision, pattern recognition, and machine learning techniques. As there is no constrained capturing in the real world, the captured documents suffer from illumination variation, viewpoint variation, highly variable scale/resolution, background clutter, occlusion, and non-rigid deformations e.g., folds and crumples. Document registration is a problem where the image of a template document whose layout is known is registered with a test document image. Literature in camera captured document mosaicing addressed the registration of captured documents with the assumption of considerable amount of single chunk overlapping content. These methods cannot be directly applied to registration of forms, bills, and other commercial documents where the fixed content is distributed into tiny portions across the document. On the other hand, most of the existing document image registration methods work with scanned documents under affine transformation. Literature in document image retrieval addressed categorization of documents based on text, figures, etc.
However, the scalability of existing document categorization methodologies based on logo identification is very limited. This dissertation focuses on two problems (i) registration of captured documents where the overlapping content is distributed into tiny portions across the documents and (ii) categorization of captured documents into predefined logo classes that scale to large datasets using local invariant features. A novel methodology is proposed for the registration of user defined Regions Of Interest (ROI) using corresponding local features from their neighborhood. The methodology enhances prior approaches in point pattern based registration, like RANdom SAmple Consensus (RANSAC) and Thin Plate Spline-Robust Point Matching (TPS-RPM), to enable registration of cell phone and camera captured documents under non-rigid transformations. Three novel aspects are embedded into the methodology: (i) histogram based uniformly transformed correspondence estimation, (ii) clustering of points located near the ROI to select only close by regions for matching, and (iii) validation of the registration in RANSAC and TPS-RPM algorithms. Experimental results on a dataset of 480 images captured using iPhone 3GS and Logitech webcam Pro 9000 have shown an average registration accuracy of 92.75% using Scale Invariant Feature Transform (SIFT).
Robust local features for logo identification are determined empirically by comparisons among SIFT, Speeded-Up Robust Features (SURF), Hessian-Affine, Harris-Affine, and Maximally Stable Extremal Regions (MSER). Two different matching methods are presented for categorization: matching all features extracted from the query document as a single set and a segment-wise matching of query document features using segmentation achieved by grouping area under intersecting dense local affine covariant regions. The later approach not only gives an approximate location of predicted logo classes in the query document but also helps to increase the prediction accuracies. In order to facilitate scalability to large data sets, inverted indexing of logo class features has been incorporated in both approaches. Experimental results on a dataset of real camera captured documents have shown a peak 13.25% increase in the F–measure accuracy using the later approach as compared to the former.
njit-etd2012-060 (112 pages ~ 10,487 KB pdf)
Please complete this Feedback Form to inform us about your experience using this website. It will assist us in better serving your information needs in the future. Thank You!
Created August 29, 2012