基于深度学习的电力设备铭牌识别  

Electricity equipment nameplate recognition based on deep learning

作　　者：陈晓龙[1] 陈显龙[1] 袁建平 高宇豆 张加其 CHEN Xiao-long;CHEN Xian-long;YUAN Jian-ping;GAO Yu-dou;ZHANG Jia-qi(Beijing Forever Technology Co., Ltd., Beijing 100011, China;School of Control and Computer Engineering,North China Electric Power University,Beijing 102206,China)

机构地区：[1]北京恒华伟业科技股份有限公司,北京100011 [2]华北电力大学控制与计算机工程学院,北京102206

出　　处：《广西大学学报（自然科学版）》2018年第6期2216-2226,共11页Journal of Guangxi University（Natural Science Edition）

基　　金：北京市科技计划课题(Z171100001217006)

摘　　要：为了获取铭牌图像中的基本参数信息,提出一种基于深度学习的端到端文本识别模型TDRN(Text Detection and Recognition Network)。模型避免了图像裁剪和字符分割,将文本看作一个序列,使用BLSTM(Bidirectional Long Short-term Memory)来获取上下文关系。同时,将文本检测和文本识别整合在同一个网络中共同训练,共享卷积层,以提高整体性能,在文本识别中还引入了注意力机制。模型在公共场景文本数据集SVT(Street View Text)上测试表现良好,F值为68. 69%,高于一般的端到端文本识别模型。与传统铭牌识别方法相比,TDRN准确率更高,鲁棒性更强,能适应复杂的电力场景变化。In order to obtain the basic parameter information in the nameplate image,an end-to-end text recognition model TDRN(Text Detection and Recognition Network)based on deep learning is proposed.The model can avoid image from cropping and character segmentation while text is treated as a sequence and BLSTM(Bidirectional Long Short-term Memory)is used to capture context.Meanwhile,text detection and text recognition are integrated in same network to be trained together and share the convolution layers to improve overall performance.Additionally,attention mechanism is introduced in text recognition.The model performs well on the public scene text dataset SVT(Street View Text)with an F value of 68.69%,which is higher than that of the general end-to-end text recognition models.Compared with the traditional nameplate recognition methods,the TDRN has higher accuracy and stronger robustness,which is able to adapt to the complicated and changing electricity scenes.

关 键 词：深度学习 文本识别 文本检测 铭牌识别 电力设备 中文文本 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]