Κατεύθυνση / ειδίκευση Επεξεργασία-Μάθηση Σήματος και Πληροφορίας (ΕΜΠ)
Πληροφορική

2020-03-30

2020

Kotsovilis Vlasios

Γάτος Βασίλειος, Ερευνητής A, Ινστιτούτο Πληροφορικής και Τηλεπικοινωνιών, ΕΚΕΦΕ Δημόκριτος

Αναγνώριση Γλώσσας Χειρόγραφων Εγγράφων με Χρήση LBP και SIFT Χαρακτηριστικών

English
Greek

Language Identification for Handwritten Documents using LBP and SIFT Features

Language identification for handwritten document images is a document analysis problem in which languages correspond to a set of graphical representations used to express a particular system of writing. Each language corresponds to unique features not only concerning the physical form, but also the writing style. Texture of an image is a unique feature that can be used to identify the language of a document image and can be defined as a repeating pattern of pixels in a structured way. In order to extract texture-based features, Local Binary Patterns (LBP) are used, which are simple in implementation and provide robustness to changes in the intensity values of image’s pixels. LBP characterizes image patches using binary codes which encode the relationship between the central pixel and its neighbours. On the other hand, gradient-based features, such as Scale Invariant Feature Transform (SIFT) descriptors, describe visual features on local regions of handwritings without the need for segmentation. Particularly, SIFT is a keypoints detection algorithm which detects local changes in the intensity of pixels in images. It also provides a sufficient number of keypoints for an in-depth use. In this thesis we present a system for automatic language identification in handwritten document images, without applying any segmentation step. Language identification can be viewed as a problem of classification in which each language represents a class. We encode text structures using texture, scale and rotation invariant descriptors derived from LBP and SIFT features respectively. LBP and SIFT features are used in experiments independently in order to extract the features from document images. Identification of language is accomplished by using K Nearest Neighbour (KNN), Naive Bayes Nearest-Neighbour (NBNN) and Local NBNN classifiers. Classification of test documents is based on the distance from features of training documents. The experiments for the evaluation of the system are performed on handwritten document images written in French, German, Greek and English languages and are part of a public dataset which contains 208 documents from 26 writers and has been used in several writer identification competitions. This thesis includes detailed results of all the above methods and it is demonstrated that language classification accuracy can reach a percentage of over 85%.

Technology - Computer science

LBP,SIFT,handwritten documents,language identification,local ΝΒΝΝ

Yes

4

Yes

22

59

https://pergamos.lib.uoa.gr/uoa/dl/object/2900014/file.pdf

https://pergamos.lib.uoa.gr/uoa/dl/object/2900014