Text Line Based Text Detection In Natural Scenes

Recent years, The internet and information technology are developing very quickly,and they have become an essential part of our life. Text, as a carrier of human thoughts and emotions, conveys high level semantics and acts an important role in many related technologies and applications. Meanwhile, text is almost ubiquitous in real world.Therefore, the relevant research on text detection and recognition have widely academic value and application value.The most of text detection methods in natural scenes are based on character detection. These methods usually need a post-processing to group character into text line or word. However, there are so many background elements are similar to characters and difficult to distinguish. Besides, the group algorithm of post-processing require high quality results of character detection. Consequently, the performance of this kind of methods is limited.In this paper, we propose to directly detect text lines to overcome these problems.For straight text in natural scenes, we propose a symmetry-based text line detection method. This method extracts the symmetry feature and self-similarity feature of text lines,and adopts Random Forest to localize text lines directly. For multi-oriented text, we propose to use Fully Convolutional Networks to detect text. We train two Fully Convolutional Networks to detect text blocks and text centroids respectively. The detected text blocks are used to generate text lines by combining MSER. Then, the character centroids within each text lines are predicted by the character centroid Fully Convolutional Networks. Finally, non-text text lines are removed by geometric constraint and intensity constraint.We evaluate these two methods on several standard benchmarks. The experimental results show our methods consistently achieves the state-of-the-art performance on these challenging datasets.