基于深度学习的智能图像处理研究  

Research on intelligent image processing based on deep learning

作　　者：刘文彬 温柏坚 高尚 郑杰生 郑颖龙 董召杰 王尧 LIU Wenbin;WEN Baijian;GAO Shang;ZHENG Jiesheng;ZHENG Yinglong;DONG Zhaojie;WANG Yao(Guangdong Electric Power Information Technology Co.Ltd.,Guangzhou 510080,China;Information Center of Guangdong Power Grid Co.Ltd.,Guangzhou 510800,China;Digital Grid Research Institute,China Southern Power Gird.,Guangzhou 510700,China)

机构地区：[1]广东电力信息科技有限公司,广州510080 [2]广东电网有限责任公司信息中心,广州510800 [3]南方电网数字电网研究院有限公司,广州510700

出　　处：《自动化与仪器仪表》2020年第8期60-63,共4页Automation & Instrumentation

基　　金：广东电网有限责任公司2016年资产管理应用升级改造研究项目(No.039900HK41160005)。

摘　　要：针对当前数显式仪表在人工采集的不便性问题,结合当前的视觉技术,提出一种基于卷积神经网络的数显仪表处理方法。为提高对数显仪表的高精度提取,首先对图像进行灰度化和二值化处理,然后对图像进行滤波和形态学处理;然后采用深度学习算法中的卷积神经网络和MOOSE算法分别对数字和小数点进行识别;最后,在搭建识别试验平台的基础上,对35 821张图像识别,识别准确率可达到95.99%,与实际识别来看,可准确识别小数点,说明本识别方法具有一定的实用价值。In view of the inconvenience of the current digital display instrument in manual acquisition,combined with the current vision technology,a digital display instrument processing method based on convolution neural network is proposed.In order to improve the high-precision extraction of digital instruments,firstly,the image is grayed and binarized,then the image is filtered and morphologically processed;then the convolution neural network and moose algorithm are used to recognize the number and decimal point respectively;finally,on the basis of building the recognition test platform,the recognition accuracy of 35 821 images can reach 95.99%.At the same time,in the actual recognition,the decimal point can be identified accurately,which shows that the recognition method obtained a certain practical value.

关 键 词：卷积神经网络 MOOSE算法 形态学 二值化 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]