摘要
文本定位作为文本识别的基础和前提,对图像深层信息的理解至关重要。针对自然场景下的文本定位受光照、复杂背景等因素影响较大的问题,提出了一种基于多方向边缘检测和自适应特征融合的自然场景文本定位方法。该方法首先将自然场景图像进行三通道八方向的边缘检测;然后通过启发式规则对得到的边缘图像进行过滤从而提取出备选文本域,进而对备选文本域进行自适应权值的HOG-LBP特征提取与融合;最后采用支持向量机进行特征分类学习,实现文本定位。实验结果表明,该方法能准确定位自然场景图片的文本区域,对光照和复杂背景具有较强的鲁棒性。
As the basis and premise of text recognition, text localization has an important influence on the analysis of images. Since the text localization in natural scene can be effected by illumination and the complex backgrounds signifi- cantly, we proposed a text localization method based on edge detection and features fusion. The method began with edge detection from three channels and eight directions, and then we filtered the detected edge images with heuristic rules to extract candidate text regions. On top of that, the HOG-LBP features were extracted and fused by adaptive weights. Fi- nally,we applied support vector machine （SVM） to classify the candidate regions and realized text localization. Experi- mental results indicate that the proposed method can locate the text region accurately in natural scene images while re- ducing the influence of illumination and complex backgrounds effectively.
出处
《计算机科学》
CSCD
北大核心
2017年第9期300-303,314,共5页
Computer Science
基金
国家自然科学基金项目(61673244)
高等学校博士学科点专项科研基金资助课题(20130131110038)资助
关键词
自然场景
文本定位
边缘检测
特征融合
Natural scene,Text localization,Edge detection,Feature fusion
