HMM Based Approach for Handwritten Arabic Word Recognition Using the IFN/ENIT- Database

An offline recognition system for Arabic handwritten words is presented. The recognition system is based on a semi-continuous 1-dimensional HMM. From each binary word image normalization parameters were estimated. First height, length, and baseline skew are normalized, then features are collected using a sliding window approach. This paper presents these methods in more detail. Some parameters were modified and the consequent effect on the recognition results are discussed. Significant tests were performed using the new IFN/ENIT - database of handwritten Arabic words. The comprehensive database consists of 26459 Arabic words (Tunisian town/village names) handwritten by 411 different writers and is free for non-commercial research. In the performed tests we achieved maximal recognition rates of about 89% on a word level.

• Type: Conference
• Year: 2003
• Where: ICDAR
• Authors: Mario Pechwitz, Volker Märgner
• Keywords: Arabic Handwritten Words, Document Analysis, Handwritten Arabic Words, ICDAR 2003, Semi-continuous 1-dimensional Hmm