基于亮度分级和方向密度的无监督文本定位

提出一种基于RGB亮度分级和方向密度的自然场景无监督文本定位方法，该方法基于场景文本通常与局部背景有较大的对比度这一特性，分别在R、G、B三个颜色层进行亮度分级，以降低背景复杂性；然后，利用文字笔画的显著方向性，以方向密度为依据进行文本区域粗定位；再进一步利用SVM多类分类器实现文本区域精确判别。新方法克服了一般无监督方法颜色聚类数目选定困难的问题，限制了候选区域的种类，从而降低了SVM分类器的训练难度，具有较高的准确性和鲁棒性。

Method for unsupervised text location based on brightness grading and direction density

A method for unsupervised text location based on brightness grading and direction density was proposed, which was according to the fact that text in scenes generally has strong contrast with local background. Brightness grading was made in R, G, B color layers separately to decrease the complexity of the background. After that, by using the characteristic of obvious directionality of text strokes, a rough text location was carried out according to direction density. And then, precise discrimination was implemented with a SVM multi-class classifier. The mentioned method overcame the difficulty to choose color clustering number in common unsupervised ways, and constrained the types of candidate areas. Hence the difficulty of training SVM classifier was reduced. Those made the new method had higher accuracy and robustness.