基于笔画角度变换和宽度特征的自然场景文本检测

Text detection based on stroke angle conversion and stroke width features in natural scene

针对光照不均和背景复杂度所导致的自然场景文本检测中文本的漏检和错检现象,提出一种基于笔画角度变换和宽度特征的自然场景文本检测方法。分析发现与非文本相比,文本具有较稳定的笔画角度变换次数和笔画宽度,针对这两个特性提出笔画外边界优劣角变换次数和增强笔画支持像素面积比两种特征。前者分段统计笔画外轮廓角度变换次数;后者计算笔画宽度稳定区域在笔画总面积的占比,用来分别反映笔画角度和宽度变化稳定特性。为降低文本漏检率,采用多通道最大稳定极值区域(maximally stable extremal regions,MSER)检测,合并所有候选区域,提取候选区域的笔画特征和纹理特征,利用支持向量机完成文本和非文本区域分类。在ICDAR2015数据库上,算法的精确率和召回率分别达到79. 3%和72. 8%,并在一定程度上解决了光照不均和复杂背景的问题。

In order to reduce the missing detection and misclassification of text caused by uneven illumination and background complexity in text detection of natural scenes,this paper presented a natural scene text detection method based on stroke angle transformation and width features.Compared to non-text,the text has a more stable performance of stroke outline angle conversion times and stroke width.Therefore,this paper proposed methods of extracting the number of transformations of the outer corner of the stroke and the enhancement of the pixel area ratio of the stroke support.In order to extract the characteristics of angular conversion,it used the method of outer contour segmentation to calculate the number of conversion times.In order to extract the strokes width characteristics,it calculated the proportion of the width stable area in the total strokes area.To reduce rate of the text missing detection,it used multi-channel MSER to detect text candidate area.Candidate areas in all channels were merged to extract the stroke and texture features.It also adopted support vector machines combined with features to classify text and non-text.The simulations show that the accuracy and recall rate of the algorithm were 79.3%and 72.8%in the ICDAR2015 database,respectively.Moreover,it solves the problem of uneven illumination and complex background to some extent.