摘要
当前电气铭牌识别效果差,无法进行工程应用.为解决电气铭牌信息识别,提出1种基于工程方法和深度学习相结合的铭牌文本信息识别ResNet50_k模型.将电气铭牌识别分为2部分:不可变区域识别和可变信息区域.针对电气铭牌可变区域的文本提取和信息识别.首先,使用变动区域位置信息对变动区域经进行获取;其次,使用K-menas聚类算法和投影法对铭牌可变区域进行分割;最后,利用Keras深度学习框架搭建残差网络模型.模型经过对3823类符的识别训练,验证准确率高达97.6%.与Tesseract OCR识别方法相比,ResNet50k效果更好.在对自然场景下拍摄电气铭牌识别中,模型表现良好,能够适应复杂的电力场环境.
The current electrical equipment nameplate(EEN) has a poor recognition effect and cannot be used for engineering applications. In order to solve the recognition of electrical nameplate information, this paper proposes an EEN text information recognition ResNetk model based on a combination of engineering methods and deep learning methods. It divides the EEN identification into two parts: the immutable area identification and the variable information area. This paper aims at text extraction and information recognition of the variable area of EEN. First, the change region is acquired by using its position information. Secondly, the K-means clustering algorithm and projection method are used to segment the variable region of the EEN. Finally, a Keras deep learning framework is used to build a residual network model. After recognition training on 3823 characters, the accuracy rate is as high as 97.6%. Compared with the Tesseract OCR recognition method, ResNet50k performs better. In the identification of electrical nameplates taken in natural scenes, the model performs well and can adapt to complex power field environments.
作者
石煌雄
胡洋
蒋作
潘文林
杨凡
张瑞祥
SHI Huang-xiong;HU Yang;JIANG Zuo;PAN Wen-lin;YANG Fan;ZHANG Rui-xiang(School of Electrical and Information Technology,Yunnan Minzu University,Kunming 650500,China;School of Mathematics and Computer Science,Yunnan Minzu University,Kunming 650500,China)
出处
《云南民族大学学报（自然科学版）》
CAS
2020年第4期350-355,共6页
Journal of Yunnan Minzu University:Natural Sciences Edition
基金
国家自然科学基金(61866040).
关键词
变动区域
字符分割
铭牌识别
残差网络
variable region
character segmentation
EEN recognition
residual network
