基于区域特征与支持向量机的场景文字定位算法

Scene Text Localization Algorithm Based on Region Feature and SVM

场景文字定位与识别问题近年来受到极大关注,本文针对定位场景图片中的文字提出一种区域与连通域结合的方法。首先采用最大稳定极值区域(MSER)与笔画宽度变换(SWT)相结合的方法来提取图片的候选文字连通域,然后运用启发规则对候选连通域初筛,最后提取候选连通域的方向梯度直方图特征(HOG)和均匀局部二值模式(LBP)特征,输入支持向量机(SVM),判别是否为文字区域。实验结果表明,本文算法具有较好的文字定位效果,并且召回率较高。

Scene text localization and recognition problem has received great attention in recent years. This paper proposes a method based on the combination of region and connected domain for the character of the scene. First, the algorithm combines the maximally stable extremal regions （MSER） with stroke width transform （SWT） to extract candidate text connected components. Then, part of the non-text region is filtered by using the heuristic rules of the candidate components. At last, the histogram of oriented gradient （HOG） and the uniform local binary patterns（LBP） of the candidate text connected components are extracted to be the input of SVM classifier. In that way, text components and no-text components can be distinguished. The experimental resuits show that the algorithm has good performance in text localization, and the recall rate is high.