Detection and Segmentation Text from Natural Scene Images Based on Graph Model
Xiaopei Liu, Zhaoyang Lu, Jing Li, Wei Jiang · 2014
This paper presents a ne w scheme for character detection and segmentation f rom natural scene images. In the detection stage, stroke edge is employed to detect possible text regions, and some geometrical features are used to filter out obvious non-text regions. Moreover, in order to combine unary properties with pairwise features into one framework, a graph model of candidate text regions is set up, and the graph cut algor ithm is utilized to classify candidate text regions as text or non-text. As for segmentation, a two-step technique for scene text segmentation is proposed. Firstly, the K-Means cluster algorithm is employed in color RGB and HSI color space respectively, and the better result is selected as initial segmentation. Then in minimum energy framework, graph cut is employed for re-labeling verification. Experimental results show the satisfactory performance of the proposed methods. Key-Words: - scene text detection text segmentation stroke width Hog feature Graph model