摘要
老式水表的人工抄表模式需要耗费大量的人工和时间成本。而当前计算机算力飞速增长,深度学习理论取得不断突破,利用神经网络和深度学习理论来实现水表读数的自动识别成为可能。为提高水表读数识别率,文章通过偏转和加噪实现数据集扩增,采用灰度化、二值化等操作对数据集进行预处理,在Tensor Flow框架下搭建卷积神经网络,选取3×3的卷积核组成三层卷积神经网络。试验结果表明,该方法的单个字符识别准确率能够达到99%,水表整体识别率稳定在97%。
The manual reading mode of old water meters requires a lot of labor and time costs. At present, computer computing power is increasing rapidly, and deep learning theory has made continuous breakthroughs. It is possible to use neural network and deep learning theories to realize automatic recognition of water meter readings. In order to improve the recognition rate of water meter readings, this article used deflection and noise addition to achieve data set amplification, used grayscale, binarization and other operations to preprocess the data set, built a convolutional neural network under the TensorFlow framework, and selected 3×3 The convolution kernel composed a three-layer convolutional neural network. The test results show that the single character recognition accuracy rate of this method can be stabilized at 98%, and the overall water meter recognition rate is stable at 96%.
作者
韦文斐
卓豫鑫
建晓鹏
Wei Wenfei;Zhuo Yuxin;Jian Xiaopeng(College of Information Science and Engineering,Henan University of Technology;School of Artificial Intelligence and Big Data,Henan University of Technology,Henan 450001)
出处
《长江信息通信》
2021年第4期26-28,34,共4页
Changjiang Information & Communications
基金
河南工业大学2019本科教育教学改革研究与实践项目JXYJ-Z201920(计算机专业硬件课程创新研究与教学实践)
河南工业大学2019年度离散数学“线上线下混合教学模式课程”。
关键词
图像处理
数字识别
卷积神经网络
水表
预处理
water meter
image processing
digital recognition
convolutional neural network
preprocessing
