基于级联CNN的自然场景文本检测

Natural Scene Text Detection Method by Cascaded CNN

目前CNN成为计算机视觉领域,特别是目标对象检测技术的主流方法之一。自然场景中的文本信息与一般目标对象不同,目标检测算法对自然场景文本检测的鲁棒性差,检测结果中的细小文本区域容易漏检,狭长文本区域检测的完整性较差。针对这一问题,对自然场景文本信息特征分析,提出了一种基于级联CNN的自然场景文本检测方法。该方法利用检测模型尽可能地发现疑似文本区域,然后利用分类模型分类筛选出最终的文本区域。在SSD目标检测算法的基础上,设计一种适用于自然场景文本的检测模型;然后对检测模型得到的疑似文本区域使用非极大值抑制和融合操作,消除重叠检测对结果的影响;最后使用针对性训练的分类模型对得到的候选区域进行分类筛选,得到最终的检测结果。该方法在数据集ICDAR2013上的召回率、准确率和F值分别为0.77、0.81和0.79,对于自然场景图像的文本检测有着较强的鲁棒性,能够有效地检测到图中细小的文本区域,明显改善狭长文本区域检测不全的情况。

CNN has become one of the mainly methods in the field of computer vision,especially in the field of targets detection.Compared of the objects in targets detection,the text in natural scene has lots of differences,so the method of targets detection has low robustness when using for text detection.And the text detection results of targets detection methods are poor,because the small text is usually dropped and narrow text area is incomplete.For this problem,the text feature in the natural scene is analyzed,and a text detection method by cascaded CNN is proposed,which uses a detection model to find more text area as far as possible and uses a classification model to screen out final result.Firstly,a detection method on the basis of SSD is proposed,which is applicable for natural scene.Then,in order to eliminate the effect of overlapping detection,the method of non-maximum suppression and regional integration is used.Finally,the candidate region are classified by ResNet classification model to receive finally results.The proposed natural scene text detection method achieves 0.77,0.81,and 0.79 in recall rate,precision rate,and F-score on the ICDAR 2013 database,respectively,and the method is robust for natural scene images,which can effectively detect the small text area and confirm the completeness of narrow text area.The experimental results show the effectiveness of the proposed method.