自然场景中文本定位方法研究

Research on Text Location Methods in Natural Scenes

自然场景中的文本信息非常丰富,然而由于场景中复杂的环境背景以及拍摄角度等因素的干扰,会导致文本定位困难。针对上述问题提出一种结合最大极值稳定区域和笔画宽度变化特征的自然场景文本定位方法:首先使用MSER对图像进行文本检测,然后对图像进行边缘化处理,并优化笔画宽度变换特征的提取,使用形态学操作去除噪点、填充间隙优化边缘图像并沿边缘像素点的梯度方向计算笔画宽度值,根据笔画宽度变化特征等获得连通域,最后根据规则过滤非文本域并合并连通域。使用ICDAR 2003文本定位竞赛数据集测试,经计算该方法有较高的正确率和召回率。通过实验结果证实了方法的有效性,可以较为准确地定位出自然场景中文本。

Text information in natural scenes is very abundant,while environmental background in natural scenes,shooting angle and other factors will lead to difficulties in locating the text. This paper proposes a method combined of maximum stability extreme region and SWT to locate text in natural scenes. Firstly,MSER is used to detect text of images and text domain is obtained.Then the image is marginalized,and the extraction process of stroke width transformation feature is optimized. The morphological operation is conducted to remove the noise,fill the gap to optimize the effect of edge image. Stroke width value is calculated along the gradient direction of the edge pixel,and then calculate the CCs. Finally non text domain is filtered according to the rules and CCs are connected. The experimental results show that the accuracy and the recall rate of the method are high. It proves that method is suitable for the text localization in the natural scenes.