Local Patterns for Offline Arabic Handwritten Recognition

Abstract

Offline recognition of Arabic handwritten text is a challenging problem due to the cursive nature of the language and high inters and intra variability in the way of writing. Majority of the existing approaches are based on structural and statistical features and are constrained for a specific task with vast amount of pre-processing steps. In this paper, we explore the performance of local features for unconstrained offline Arabic text recognition with no prior assumptions or pre-processing steps. Our approach is based on local SIFT features. To capture important information and remove any redundancy, we apply a fisher encoding algorithm, and a dimensionality reduction approach, principle component analysis (PCA). The resulted features are combined with a contemporary support vector machine (SVM) classifier and tested on a dataset of 12 different classes. There has been great improvements in recall and precision values in comparison with that of SIFT features alone or with that of SIFT features and other encoding algorithms, with more that 35% improvements when tested with 5-fold cross-validation test.