基于笔画生长的自然场景艺术文字检测  

Wordart Detection in Natural Scene Based on Stroke Growing

作　　者：梁添才 罗攀峰[1,2,3] 张永[2,4] 聂芸芸 

机构地区：[1]广州广电运通金融电子股份有限公司广电运通研究院,广东广州510663 [2]广东省货币识别企业重点实验室,广东广州510663 [3]清华大学机械工程学院,北京100084 [4]中山大学信息科学与技术学院,广东广州510006

出　　处：《计算机仿真》2015年第8期284-288,共5页Computer Simulation

基　　金：广东省战略性新兴产业核心技术攻关"面向金融电子装备业的货币智能分拣关键技术研发"(2011A091101002);广东省重大科技专项项目"基于物联网的货币流通安全平台研究和开发"(2012A080104005)

摘　　要：自然场景图像中的文字信息属于高层语义信息,通常与图像内容紧密相关,是理解图像场景内容的重要线索。由于图像中的文字通常具有复杂的字体变化,使得现有的文字检测技术难以获得较好的检测结果。为了解决场景文字检测受到字体自身笔画属性的影响,提出一种针对笔画宽度不均匀的艺术字体的检测方法。首先采用笔画宽度变换算法沿垂直于图像边缘的方向搜索平行边缘对,确定笔画候选位置;其次在具有强平行特征的边缘对内侧采样笔画前景种子像素,在边缘对外侧采样背景种子像素。采用核密度估计方法计算图像范围内任意像素属于笔画像素的似然函数;最后以该似然函数为权重函数,采用基于颜色特征的区域生长法,将笔画覆盖范围由具有强边缘平行特征的前景种子像素蔓延至完整笔画区域,从而实现笔画区域的精确提取。实验表明上述方法具有快速、有效、稳定的特点,尤其适用于自然场景中笔画宽度不均匀的艺术字体的检测。This paper presents a novel algorithm for segmenting Wordart form natural scene. The method is divid- ed into three steps. In the first step, the Stroke Width Transform （SWT） algorithm is used to define the strong strokes regions. The foreground seeds and background seeds of strokes are then identified according to these strong strokes regions. In the second step, a weight function is set to the likelihood that a pixel belongs to the strokes fore- ground. This likelihood is obtained from the foreground seeds and background seeds in Luv color space. The proba- bility density function is the color PDF of image space, which is obtained via the fast kernel density estimation. In the third step, a region growing algorithm is proposed. The process is iterative, in which the strokes segmentation is reliably computed automatically no additional user effort is required. Experiments show that the proposed method is superior to existing methods in terms of efficiency, effectiveness, and robustness, and can be widely applied in seg- menting of Wordart form natural scene.

关 键 词：文字分割 笔画宽度变换 区域生长 艺术字检测 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]