基于启发式规则和SVM的自然场景中文文本定位

摘要

To realize the Chinese text location in the natural scene,a level positioning method combined with heuristic rules filtering and SVM scientific classification is designed. Firstly, the Maximally Stable Extremal Regions algorithm is adopted to achieve stroke amalgamation and consequently the heuristic rules are made based on the characteristics of Chinese characters to filter non-text regions. By ellipse fitting in the candidate character zone, eccentricity ratio of ellipse is taken as decision rule of text, and finally given the extracted HOG features, SVM is used to do accurate classification to realize text location. It is shown in the experiment that proposed method in the paper gets good test location effect in the complex natural scene.%为了实现自然场景下的中文文本定位，设计实现了一种启发式规则过滤和SVM精确分类的层次定位方法。首先通过最大稳定极值算法提取区域，对于汉字笔画分离的问题，用形态学运算进行笔画融合。再根据汉字的特点，设计启发式规则过滤非文本区域，其中通过候选字符区域的椭圆拟合，引入椭圆的偏心率作为文本判别规则。最后提取HOG特征，通过SVM精确分类实现文本定位。实验证明本文方法在复杂的自然场景下取得了良好的文本定位效果。