A SWT Verified Method of Natural Scene Text Location
Jian Huang, Xiaopei Liu, Qian Zhao · 2016
Text location from natural scene images is very challenging due to variations in text fonts, color, alignment and illumination. In order to deal with this problem, a novel method of scene text location is proposed. Firstly, gradient difference is introduced to detect text regions, because it can capture the texture various in the local regions. Secondly, a stroke width based verification step is employed to wipe off false alarms, and stroke width of candidate characters is computed by SWT (Stroke Width Transform). The proposed method is evaluated on ICDAR2003 location contest dataset, and the experimental results show it outperforms methods with similar ways.