Detection and Recognition of Arabic Text in Video Frames

The authors have developed an end-to-end system for Arabic text recognition in video frames. The end-to-end system consists of the steps for text-line detection, word segmentation and word recognition. In order to achieve high text recognition accuracy we propose a new scheme of integrated text detection-recognition scheme, where the true text-lines are detected with as higher recall rate as possible and the false words in the false lines are rejected in the successive word recognition step. We reported a recognition based transition frame detection of Arabic news captions in single channel video images [9]. In this paper the recognition system is integrated with n-gram language model and extended to text detection/recognition of multi-channel video images. The multi-channel, multi-font performance of the system is experimentally evaluated using AcTiV-D and AcTiV-R dataset. The multi-channel text detection performance for three channels, France24, Russia Today and TunisiaNat1 is 91.29% in F-measure. The multi-channel, multi-font character recognition performance for these channels is 94.84% in F-measure.