基于DenseNet自然场景文本检测  

Natural scene text detection based on DenseNet

作　　者：刘会江 曾浩[1] 陈阳 LIU Hui-jiang;ZENG Hao;CHEN Yang(College of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《计算机工程与设计》2020年第8期2201-2206,共6页Computer Engineering and Design

基　　金：重庆市教委自然科学基金项目(KJ1500407)。

摘　　要：针对自然场景文本图片背景的多样性、多变性、复杂性等问题,提出一种基于密集连接网络(DenseNet)的DenseText模型。使用改进的DenseNet网络进行提取特征;为更好适应文本图片字体较长的特点,采用长方形卷积核代替传统网络的正方形卷卷积核;区别于传统的NMS后处理算法,使用Soft-NMS算法进行优化。识别框架上,采用CRNN网络进行识别,形成一个端到端的自然场景文本检测与识别的一体化网络框架。实验结果表明,该网络模型在ICDAR13数据集上取得了令人信服的结果,提高了检测结果的准确率,降低了假阳性,recall达到0.85,F为0.88。To solve the problems of diversity,variability and complexity of text background pictures in natural scenes,a DenseText model based on DenseNet(dense connected network)was proposed.The improved DenseNet(densely connected network)was used for features extraction.To better adapt to the longer features of text image fonts,rectangular convolution kernels were used instead of traditional network square volume convolution kernels.The traditional NMS post-processing algorithm was optimized using the Soft-NMS algorithm.On the frame of recognition,CRNN was used to identify,and an end-to-end natural scene text detection and recognition framework was formed.Experimental results show that the DenseText model achieves convincing results on the ICDAR13 dataset,which improves the accuracy of the test results and reduces the false positives.The recall is 0.85 and F is 0.88.

关 键 词：自然场景 文本检测 文本识别 长卷积核 密集连接网络 软非极大值抑制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]