Offline Sub- Word Handwritten Recognition for Arabic Historical Manuscripts

Abstract

In this thesis, we address Arabic offline handwritten recognition for historical documents. The Automation of the handwritten recognition has many applications, such as zip coding, forms processing, indexing and retrieving historical manuscripts and so on. Recognition for Arabic handwritten script lags far compared to other languages such as Latin, and Chinese texts. The challenges for Arabic language raise from its nature such as overlapping characters, cursive texts, and lack of benchmark databases. In this work, we introduced new techniques for different phases of the offline Arabic handwritten recognition. First it addresses the recognition of the Arabic handwritten for historical manuscripts not contemporary scripts. In addition, the feature selection algorithm is presented. This work aims to select appropriate features and remove irrelevant ones. The relevant features are those which enhance the results and give a higher success rates. We depend on probabilistic classifier not statistically such as HMM. Naive Bayesian classifier is used for training and classification. The presented work was applied and tested on private database, which was collected from historical manuscripts. A competitive recognition rates were achieved. The results show that applying feature selection prior classification gives haghier success rates than classification without feature selection