基于深度全卷积神经网络的文字区域定位方法  

Text area location method based on depth full convolutional neural network

作　　者：骆遥[1] 

机构地区：[1]同济大学测绘与地理信息学院,上海200092

出　　处：《无线互联科技》2016年第23期43-44,共2页Wireless Internet Technology

摘　　要：近年来,深度学习模型在各种计算机视觉方面都展现出了远远优于传统方法的性能,在自然场景中的文字区域定位问题中引入深度学习方法无疑也是大势所趋。文章提出了一种基于深度全卷积网络方法的文字区域定位方法,实现了端到端的训练、检测,使得训练更为有效,检测过程更加高效。最终文中方法在ICDAR 2015数据集上对比基于MSER等的传统方法有了很大提升,达到了86.57%的查准率和82.1%的召回率。Deep learning has drawn lots of attention recently due to its powerful ability in both computer vision and voice field. Introducing depth learning method in text area localization problem of natural scene undoubtedly is the trend. In this paper we proposed a new method based on deep fully convolutional networks for neural scene text localization task which is an end-to-end method.The method we proposed makes the training and detection much more efficient compared with the traditional method such as MSER method. Finally we achieved 86.57% precision and 82.1% recall in ICDAR 2015 data set.

关 键 词：深度全卷积网络 自然场景文字区域定位 图像区域分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]