Περίληψη
Τα σύγχρονα συστήματα αναγνώρισης γραφέα χρησιμοποιούν μια ποικιλία διαφορετικών χαρακτηριστικών και τεχνικών για να προσδιορίσουν τον συγγραφέα του χειρόγραφου κειμένου. Σε αυτή την διατριβή παρουσιάζονται διάφορα κατευθυντικά χαρακτηριστικά καθώς και συνδυασμοί κατευθυντικών χαρακτηριστικών με χαρακτηριστικά που βασίζονται σε μοντέλα. Συγκεκριμένα, επιχειρούνται αρκετές βελτιώσεις ενός στατιστικού, κατευθυντικού χαρακτηριστικού, του edge hinge distribution. Τα νέα χαρακτηριστικά που παρουσιάζονται είναιτο Skeleton Hinge Distribution, το Weighted Skeleton Hinge Distribution, το Quantized SkeletonHinge Distribution, το Directional Stroke Run Length Distribution και το Edge Skeleton HingeCombination . Επιπλέον, διερευνάται ο συνδυασμός του Skeleton Hinge Distribution με ένα χαρακτηριστικό που βασίζετε σε μοντέλα. Νέες συνεισφορές που σχετίζονται με την προεπεξεργασία των εικόνων εγγράφων αλλα και την εξαγωγή πολύτιμων χαρακτηριστικών του κειμένου. Ειδικότερα, παρουσιάζονται δύο τεχνικές ...
Όλα τα τεκμήρια στο ΕΑΔΔ προστατεύονται από πνευματικά δικαιώματα.
The state-of-the-art writer identification systems use various features and techniques to identify the writer of the handwritten text. In this work, several directional features and combinations of directional with model-based features are presented. Specifically, several improvements of a statistical, directional feature, the edge hinge distribution, are attempted in novel contributions as the Skeleton Hinge Distribution, the Weighted Skeleton Hinge Distribution, the Quantized Skeleton Hinge Distribution, the Directional Stroke Run Length Distribution and the Edge Skeleton Hinge combination. Furthermore, the Skeleton Hinge Distribution feature with a model-based feature is explored, based on a codebook of graphemes. Novel contributions related to the preprocessing of the document images and the extraction of valuable characteristics are presented. More specifically, two techniques are presented for MainBody Size estimation, a characteristic with application in a broad range of documen ...
The state-of-the-art writer identification systems use various features and techniques to identify the writer of the handwritten text. In this work, several directional features and combinations of directional with model-based features are presented. Specifically, several improvements of a statistical, directional feature, the edge hinge distribution, are attempted in novel contributions as the Skeleton Hinge Distribution, the Weighted Skeleton Hinge Distribution, the Quantized Skeleton Hinge Distribution, the Directional Stroke Run Length Distribution and the Edge Skeleton Hinge combination. Furthermore, the Skeleton Hinge Distribution feature with a model-based feature is explored, based on a codebook of graphemes. Novel contributions related to the preprocessing of the document images and the extraction of valuable characteristics are presented. More specifically, two techniques are presented for MainBody Size estimation, a characteristic with application in a broad range of document image analysis fields. One measures Main Body size directly, while the other does an estimation for the baselines first. Both methods are segmentation free. A collection of 10 printed document images and a collection of handwritten text were used for the presented experimental results. Furthermore, a technique for text localization is presented that takes advantage of the fact that text should present some contrast in comparison with the background, to be distinguished by the humaneye. A procedure of binarization is applied to create appropriate images for text detection. The connected components of the image are extracted, and some heuristic rules are applied to identify areas containing text. For the evaluation, the Firemaker Database and the ICDAR 2017 writer identification competition dataset were used. A plethora of matching techniques were considered for Skeleton Hinge distribution, including nearest neighbour classifier, K-means, Hierarchical Cluster Tree, k-nearest neighbours and Support Vector Machines. The skeleton hinge distribution achieved an accuracy of 90,8%, while the combination of this method with the codebook of graphemes reached 96%. The Weighted Skeleton Hinge Distribution achieved an accuracy of 91.2%. The QuantizedSkeleton Hinge Distribution achieved an accuracy of 92.4%. The Directional Stroke Run LengthDistribution achieved an accuracy of 91.2%, and finally, the Edge Skeleton Hinge combination technique achieved an accuracy of 90,2%.Keywords: Writer Identification, Edge-Hinge Distribution, Skeleton-Hinge Distribution, Codebook of Graphemes
περισσότερα