Paper

An Enhanced Bag-of-Features Framework for Arabic Handwritten Sub-words and Digits Recognition


Authors:
Mohammed O. Assayony; Sabri A. Mahmoud
Abstract
In recent years, feature learning approaches have gained substantial interest and are successfully applied to challenging problems in facial recognition, visual object retrieval and classification, document image analysis and, recently, in handwriting recognition. In this paper, we present a feature learning framework for Arabic handwritten text recognition based on the Bag-of-Feature (BoF) paradigm. Utilizing the characteristics of handwritten text, we developed two novel versions of SIFT that are discriminative and computationally efficient with half the size of the original SIFT descriptors. To evaluate the quality of the features learned by the framework and the efficiency of the proposed versions of SIFT, we conducted extensive experimental work on two Arabic handwritten text datasets, viz. the non-touching Arabic Indian digit and Arabic sub-words datasets of CENPARMI Bank check database. Our framework achieves state-of-the-art accuracies on both datasets. The recognition performance and the computational efficiency are the result of utilizing the unique properties of the handwritten text.
Keywords
Feature Learning; Bag-of-Features; Arabic Handwriting Recognition; SIFT
StartPage
27
EndPage
38
Doi
Download | Back to Issue| Archive