Scene text extraction based on symmetrical edge-point pair detection of character stroke
Teng Liu, Liang Zhou · 2017
Scene text extraction is always a challenging task owing to its usual disturbing factors such as complex image backgrounds and various text behaviors (sizes, colors, styles and alignments). This paper proposes a scene text extraction approach based on the novel concept of `symmetrical edge-point pairs' (`point-pair'), which is adopted to describe the sizes, directions and brightness information of text character strokes. First of all, `point-pair' samples from both symmetrical edges are obtained based on the Laplacian of a Gaussian (LoG) method, according to the capturing capacity towards the edge light-dark trends of the second-order edge detection. Secondly, a series of stroke adaptive search windows, which is formed by the distribution of the samples, are used to detect stroke connecting regions. Thirdly, Minimum Spanning Tree (MST), pruning, and false detection elimination algorithms are employed to cluster all stroke candidate regions into one cluster. Finally, required scene texts are extracted from the candidate cluster by an adaptive segmentation threshold. Experimental results show that the proposed approach can efficiently and precisely extract the scene texts within the various disturbing factors (multi-scale, multi-direction, various brightness types of text, multi-language and complex backgrounds).