A natural scene text extraction method based on the maximum stable extremal region and stroke width transform

To extract text information effectively from natural scene image with complex background, multi-orientation perspective and multilingual languages, a scenario text extraction method based on maximum stable extremal region (MSER) and stroke width transform (SWT) is presented. The method combines the merits of MSER and SWT algorithms. It establishes text candidate regions by utilizing MSER algorithm to detect text regions, and SWT algorithm is used to calculate the text stroke width of candidate region to get its stroke width. According to the stroke width graph, the stroke-connected graph is established by using connected component labeling. Then the heuristic rules of stroke-connected graph are established to remove non-text candidate regions according to the stroke-connected graph. By using the geometrical feature analysis and the local adaptive window Otsu segmentation, the text in natural scene images can be extracted effectively. The text extraction accuracy, recall rate and comprehensive performance of this method are 0.74, 0.64 and 0.68, respectively. Simulation experiment shows that the method can achieve good robustness for complex background with multi-orientation perspective, various characters and font sizes, and it is suitable for variety of languages and fonts.