最大稳定极值区域与笔画宽度变换的自然场景文本提取方法

发布时间：2019-05-09 22:31

【摘要】：针对从背景复杂、视角多变、语言形式多样的场景图像中难以准确提取文本信息的问题,提出了一种基于最大稳定极值区域(MSER)和笔画宽度变换(SWT)场景文本提取方法。该方法结合MSER、SWT算法的优点,采用MSER算法的准确检测文字区域,建立文本候选区域,利用SWT算法计算文本候选区域笔画宽度得到候选文本区域的笔画宽度;根据笔画宽度图,利用连通域标记建立笔画宽度连通图,然后根据笔画宽度连通图,建立笔画连通图的启发性规则,删除非文本候选区域,并根据文本的几何特征分析及局部自适应窗口最大类间方差(Otsu)分割,有效提取出自然场景图像中的文本,文本提取的准确率、召回率及综合性能分别为0.74、0.64及0.68。仿真实验结果表明,在文本视角多变,字符大小、尺寸、字体各异的复杂条件下,所提方法具有较好的鲁棒性,适用于多语言和多字体混合的场景文本提取。
[Abstract]:In order to solve the problem that it is difficult to extract text information accurately from scene images with complex background, changeable visual angles and diverse language forms, a (SWT) scene text extraction method based on maximum stable extremum region (MSER) and stroke width transformation is proposed. Combined with the advantages of MSER,SWT algorithm, this method uses MSER algorithm to detect the text region accurately, establishes the text candidate region, and calculates the stroke width of the candidate text region by using SWT algorithm to obtain the stroke width of the candidate text region. According to the stroke width diagram, the stroke width connection map is established by using the connected domain mark, and then the enlightening rules of the stroke connection graph are established according to the stroke width connection graph, and the non-text candidate regions are deleted. According to the geometric feature analysis of the text and the (Otsu) segmentation of the maximum variance between classes of local adaptive windows, the text in the natural scene image is effectively extracted, and the accuracy of the text extraction is achieved. The recall rate and comprehensive performance were 0.74, 0.64 and 0.68, respectively. The simulation results show that the proposed method has good robustness under the complex conditions of variable text perspective, character size, size and font, and is suitable for multi-language and multi-font mixed scene text extraction.
【作者单位】：西安交通大学电子与信息工程学院;
【基金】：国家自然科学基金资助项目(61306111)
【分类号】：TP391.41

【相似文献】