Authors: Cunzhao Shi, Yang Zhang, Chunheng Wang, Baihua Xiao, Song Gao, Jinlong Hu

Description: We propose a robust text detection algorithm, in which Maximally Stable Extremal Regions (MSER) are used as the candidate character components [1], and tree-structured character models (TSM) [2] are used to search the missing characters and eliminate the false positives. First, two kinds of MSERs, dark-on-light and light-on-dark ones, are detected. Too large or small regions which are obviously non-text ones are eliminated at this stage. The following process is applicable for both kinds of MSERs. Then, for each kind of MSER, we use a region based classifier (detailed in [1]) to exclude some non-text MSERs and the left text candidates are grouped into text lines according to the position, size and color of each MSER. Next, to eliminate the false positives and also search the missing characters, we apply the TSM on each candidate text region. On one hand, if the detection scores of the TSM are too low, the region would be eliminated. On the other hand, if the detection scores of the TSM on the region are higher than a certain value, we enlarge the region and apply the TSM to search possible missing characters. Finally, the text lines are partitioned into words and results from the two kinds of MSERS are merged.

Reference
[1] Cunzhao shi, Chunheng Wang, Baihua Xiao, Yang Zhang, Song Gao. “Scene Text Detection Using Graph Model Built Upon Maximally Stable Extremal Regions.” Pattern Recognition Letters, Vol. 34, No. 2. (January 2013), pp. 107-116.(SCI)
[2] Cunzhao Shi, Chunheng Wang, Baihua Xiao, Yang Zhang ,Song Gao and Zhong Zhang. “Scene Text Recognition using Part-based Tree-structured Character Detections”, CVPR 2013，in press.