Text detection in Arabic news video based on MSER and RetinaNet

Abstract

In this paper, we propose a novel approach for text detection in Arabic news videos. Firstly, we apply MSER method and morphological operators (open and close) to extract candidate regions of text in image. Then, we use a deep learning method called RatinaNet. It is based in two stages. The first one aims to extract features using residual network (ResNet) and a pyramidal feature network (FPN). In the second step, we use two fully convolutional networks (FCN), one is for the classification task and the other for the bounding box regression task. For training and testing stages, we have used the AcTiVD [18] dataset. Experiments results proves the efficiency and performance of the proposed method.