Arabic text detection in video scene using Deep Learning

Abstract

Our thesis focuses on the detection of Arabic text in video scenes using computer vision techniques by comparing two approaches: a traditional method represented by the MSER (Maximally Stable Extremal Regions) algorithm, and a modern method based on deep learning using the YOLOv5 model. The MSER algorithm showed poor performance in detecting Arabic text in videos with complex backgrounds, with accuracy dropping to as low as 13%. However, in videos with simple backgrounds, its accuracy reached up to 74%, indicating a relatively good ability to detect Arabic text in less challenging scenes. On the other hand, the YOLOv5 approach proved to be highly effective across various background conditions, achieving a high accuracy rate of up to 98% in detecting Arabic text in video scenes. Nevertheless, this method requires significantly more time and effort for data preparation and model training.