Structure Context Clues for Chinese Text Detection

In this paper, we focus on Chinese text detection in a natural scene image. Different from the existing methods which usually rely on complicated features or heavy learning models, our method introduces simple structure features and context groups which are very effective for Chinese text detection. Specifically, we firstly extract connected components by using the Maximally Stable Extremal Region algorithm (MSER) in multiple channels of color space and the candidate regions are generated by combining the extraction results. Secondly, we design some simple structure features which are special for Chinese characters to filter non-text candidate regions. Some detached parts of a character are merged to form a whole character in this stage. Finally, we propose the structure context clues. Making use of the context difference between text and non-text, we further to filter the non-text regions by a context-based group filter and a Supported Vector Machine (SVM) classifier trained with Histogram of Oriented Gradients (HOG) features. Furthermore, we build our Chinese Character Street View (CCSV) dataset, on which our approach is implemented. And the experimental results demonstrate the availability of our method.