基于笔画识别的视频图片文字提取方法

通过对笔画的对称边缘特点与文字几何特征的认识,根据二阶边缘检测算子捕捉边缘点亮暗变化趋势的能力,使用高斯型拉普拉斯算子寻找"边缘点对",并构建来自笔画等窄带区域的"对称边缘点对"样本集.从样本集的分布规律中自适应地定出文字笔画搜索窗的尺度及方向.利用最小生成树算法实现由系列搜索窗得到的所有笔画子区域的关联聚类,通过剪枝、伪区域鉴别和阈值分割,将文字以行(含非水平行)或列的形式提取出来.实验表明,该方法对复杂背景下不同的语言类型、亮暗类型、文字行方向及文字尺度具有适应性,在ICDAR数据集上的查准率和查全率分别达到76%和75%.

Text Extraction Based on Stroke Recognition in Video

According to geometric features of texts and the fact that character strokes have symmetrical edges, Laplacian of a Gaussian（LoG）was employed for finding the ‘symmetrical edge-point pair’,then the ‘symmetrical edge-point pairs’ sample set was constructed,therefore the scale and orientation of the detect window were deter-mined by analyzing the sample distribution. The relational cluster of all character sub-regions was obtained by using the minimum spanning tree（MST）algorithm,then the text lines（including non-horizontal）were extracted in the form of lines or rows after pruning,false positive elimination,and threshold segmentation. Experiments show that the proposed method is capable of handling multilingual,different orientation and multi-scale images under complex background with a 76%precision rate and a 75%recall rate on ICDAR dataset.