结合MSCRs与MSERs的自然场景文本检测

发布时间：2018-03-08 19:33

本文选题：自然场景　切入点：复杂背景　出处：《中国图象图形学报》2017年02期 　论文类型：期刊论文

【摘要】：目的目前,基于MSERs(maximally stable extremal regions)的文本检测方法是自然场景图像文本检测的主流方法。但是自然场景图像中部分文本的背景复杂多变,MSERs算法无法将其准确提取出来,降低了该类方法的鲁棒性。本文针对自然场景图像文本背景复杂多变的特点,将MSCRs(maximally stable color regions)算法用于自然场景文本检测,提出一种结合MSCRs与MSERs的自然场景文本检测方法。方法首先采用MSCRs算法与MSERs算法提取候选字符区域;然后利用候选字符区域的纹理特征训练随机森林字符分类器,对候选字符区域进行分类,从而得到字符区域;最后,依据字符区域的彩色一致性和几何邻接关系对字符进行合并,得到最终文本检测结果。结果本文方法在ICDAR 2013上的召回率、准确率和F值分别为71.9%、84.1%和77.5%,相对于其他方法的召回率和F值均有所提高。结论本文方法对自然场景图像文本检测具有较强的鲁棒性,实验结果验证了本文方法的有效性。
[Abstract]:At present, based on MSERs (maximally stable extremal regions) the text detection method is the main method of natural scene text detection. But some of the text in natural scene images of complex background and MSERs algorithm can not be accurately extracted, reduces the robustness of this method. Based on the background of complex natural scene text image the characteristics of the MSCRs (maximally stable color regions) algorithm for natural scene text detection, proposed a combination of MSCRs and MSERs in the natural scene text detection method. The method first uses MSCRs algorithm and MSERs algorithm to extract the candidate character area; then using random forest texture feature training character classifier candidate character area, to classify the candidate characters area, and get the character region; finally, based on the character of the regional color consistency and geometric relations of adjacent words Agree to merge, to get the final results. The results of this text detection method in ICDAR 2013 recall rate, accuracy and F-measure were 71.9%, 84.1% and 77.5%, compared with the recall of other methods and F-measure were improved. Conclusion this method is robust to the natural scene text detection, the experimental results to verify the validity of this method.

【作者单位】：武汉大学印刷与包装系;
【基金】：国家自然科学基金项目(61601335) 国家科技支撑计划资助项目(2013BAH03B01) 国家测绘地理信息局卫星测绘技术与应用重点实验室经费资助项目(KLSMTA-2016-04) 中国博士后科学基金项目(2015M582277) 中央高校基本科研业务费专项基金项目(2042015kf0059)~~
【分类号】：TP391.41

【相似文献】