Deep Learning Approaches in Arabic OCR

Abstract

Recognition of Arabic text in printed documents or natural scenes is a hard problem compared to the same application on Latin languages. Since Arabic typography is more complicated and the difference between characters can be very subtle there becomes more steps needed in such systems we want to build. In this thesis we develop Arabic characters’ recognition systems for different tasks. The first recognition pipeline is used to recognize and transcribe Arabic licenses car plates from a video stream. The pipeline consists of a localization algorithm for extracting the license plate, and a deep neural network for recognizing characters on the plate. The neural network was trained on synthetic data and tested on a real world example manually annotated. The model for Arabic character recognition achieved 90% accuracy, while the model for Arabic numbers recognition achieved 94% accuracy. We then introduce an Arabic OCR system for recognizing Arabic text in scanned documents. The pipeline starts with a document containing lines of text, and segments them into each line alone then into sub-words using histogram projection thresholding. Since the remaining sub-systems of the pipeline are trained to recognize text of default size 18pt, we built a neural network model for predicting the size of an input sub-word, hence we can afterwards normalize this sub-word into the 18p default size. A multichannel deep neural network model is built to segment input sub-word into characters. Then a model for recognizing characters is finally used to have the final output