基于对象建议算法的自然场景文本检测

对象建议算法(object proposals)是对象检测中的常用算法,用于快速定位物体区域。根据自然场景文本的特点,将对象建议算法应用到文本检测中,并与经典的最稳定极值区域算法相结合;然后,通过贝叶斯模型融合了笔画宽度特征、视觉散度特征和边缘梯度特征,并将文本和非文本区域的区分问题转换成一个二值标记问题,通过最小化能量函数寻找最佳标记;最后,通过均值漂移聚类寻找文本区域的中心生成文本行。经实验证明,本算法在常用的自然场景文本检测数据集上速度得到了提高,并且一定程度上解决了传统最稳定极值区域算法对光照敏感的问题,获得了较高的查全率。

Scene Text detection based on Object Proposals

Object Proposals is often used in Object Detection. This paper apply object proposals to text detection and is combined with Maximally Stable Extremal Region (MSER) by analyzing the text feature in natural scene. And then it used Stroke Width,Perceptual Divergence,Histogram of Gradients at Edges and a Bayesian method for their integration. And we cast the task of separating characters from non-characters as a binary labeling problem, minimizing an energy function to find the best label. Finally we use mean-shift to find the text line. Experiments show that, this algorithm increases the detecting speed, and it solves the problems that MSER is sensitivity to blur and uneven illumination, increase the recall in a sense.