Scene text detection based on enhanced multi-channels MSER and a fast text grouping process

Scene text detection has become a popular research in computer vision and pattern recognition field in recent years because of the accurate and rich information carried by scene text. Now component-based methods have become the trend, and the detection result is largely determined by the success of filtering text-like non-text regions. The main task of this paper is to reduce the time complexity without a big fall in recall. First an enhanced multi-channels MSER model is introduced: Before extracting MSER, the image is sharpened by using the Laplacian and Gaussian blur and multi-channels is utilized, then the step of the threshold used in MSER algorithm is set to the minimum in order to get more refined MSERs. Second, two novel scene text features local contrast and boundary key points are proposed to better distinguish text regions from non-text regions. Finally, a fast text grouping algorithm is achieved which reduce the time complexity. Experiments on both ICDAR 2011 and ICDAR 2013 show that the recall of the proposed method is improved by 3%.