Feature Extraction for Cursive Language Document Images

Abstract

The efficiency of any machine learning and computer vision system depends largely on the robustness of feature extraction and selection process. In word spotting applications, many appropriate features have been proposed over the years in literature. Most of these features are extracted for Latin text but are used with Oriental script as well. Extracting features that are more specific to Oriental text is also being investigated and a lot of research is being focused on this aspect lately as well. Deep Learning has also been employed for this purpose. In this paper, we have tried investigate the performance of shape based features for Urdu script. Urdu and Arabic belong to the same family of script and both share similar set of alphabet. This means that features investigated on Urdu will give similar performance for Arabic as well as other Oriental scripts. For this paper, we have compiled results on approximately 21000 ligatures belonging to 200 unique classes taken from scanned pages of the popular Urdu series ‘Zaawiyya’. This is Higher Education Commission granted project, due to this data set is provided by them. Proposed system gives encouraging results with precision of 88.5% and recall rate of 90.8%.