Digital video libraries know a huge growth especially with video sharing tools, delinearizedTV programs and video-on-demand systems. Representing, searching, and retrieving such type of data become an important issue that requires the use of all the information available outside and within the video. Embedded text that is artificially added in a video, like places names and news subtitles, is one of the most high-level semantics used in content-based video structuring and retrieval.
In this thesis, we are interested in the detection and the recognition of Arabic embedded text in videos. The focus on Arabic text is motivated by many reasons. First, this language is used by more than half of a billion people in the world and many big and popular Arabic news channels appeared in the last two decades. Second, Arabic text has many specific properties that make its detection, and recognition in a second step, very challenging (cursive text, more different shapes than Latin or Chinese text, etc.).
The first step in this thesis is to propose a detection method of Arabic embedded text in videos that separate it from the background and precisely localize in each frame. Then, the second step consists in recognizing this text with mainly machine learning-based approaches. The final Arabic OCR (Optical Character Recognition) system must be robust to text variability (different fonts, sizes, colors, etc.) and acquisition conditions (low resolution, complex background, non uniform lighting, etc.).