Efficient recognition of machine printed Arabic text using partial segmentation and Hausdorff distance

There is an urgent need for reliable and efficient systems for off-line automatic reading of machine printed Arabic texts. A partial list of applications that may use such system includes searching and reading in scanned books and manuscript as a part of digital libraries; recognizing text on digitized maps, vehicle license plates, road signs and others. In this research we aim to contribute to the research of recognizing Arabic machine printed texts using a partial segmentation process and Hausdorff distance. The process analyses the layout of the image and segments it to words and Parts of Words (PAWs). The Stroke Width Transform (SWT) is used to calculate the size and the font in order to define a set of multi size sliding windows to search and identify characters within the given shape of a PAW. The process evaluates the similarity of the two sub images (character and sliding window) using Hausdorff distance. The top k - ranked candidates and their places within the PAW are recorded and used to generate a list of full PAWs images. In the next step elements of this list are matched to the given shape in a holistic manner. We have tested our approach using the APTI, the PATS- A01 data sets and a private collection of text images and encouraging results were obtained.