Research On Natural Scene Text Detection

The growing popularity of smart mobile terminals and the increasing convenience of Internet access have brought more extensive utilization of image-based applications, such as contentbased multimedia information retrieval, language translation, scene understanding, assistive navigation, and human-computer interaction, where natural scene texts are widely captured and shared in the form of images. As a key technology to achieve the above functions, text detection is directly related to the precision, robustness, and generalization ability of the entire imagebased application. Natural scene text images generally have a complicated background with texts not so well organized. Rich details also bring a lot of uncertainty for detection. Therefore, research on the natural scene text detection has important theoretical and practical significance.In this paper, a stepwise text detection framework is proposed. The proposed framework integrates Maximally Stable Extremal Regions(MSER), Stroke Width Transform(SWT) and GrabCut algorithm, and achieves reliable detection of scene text in a coarse-to-fine way. To be more specific, this paper implements the natural scene text detection algorithm based on MSERs, and makes use of spatial features to filter and merge the detection outputs, then gets the fast segmentation of candidate text regions. A filtering mechanism based on character stroke width feature is used to improve the reliability and robustness of the detection outputs. The filtering mechanism also combines the MSER detection algorithm and the SWT tightly. By means of a scene text extraction operation based on GrabCut algorithm, the proposed framework further removes background interference, while interaction options provided by the GrabCut algorithm can be picked to correct the wrong or missing detection situation. The efficient text extraction results can be directly used for text segmentation or recognition. The proposed framework does not use the training set, and interaction helps to further improve the performance. At the same time, the proposed framework has the characteristics of high efficiency and lightweight, and is suitable for transplantation to the mobile terminal applications.