基于深度学习的字轮式水表读数检测与识别  

READING DETECTION AND RECOGNITION OF WORD WHEEL WATER METERBASED ON DEEP LEARNING

作　　者：陈妃奋 苏健 张红梅[1] 张向利[1] Chen Feifen;Su Jian;Zhang Hongmei;Zhang Xiangli(School of Information and Communication,Guilin University of Electronic Science and Technology,Guilin 541004,Guangxi,China;Guangxi Shangshan Ruoshui Development Co.,Ltd.,Nanning 530029,Guangxi,China)

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004 [2]广西上善若水发展有限公司,广西南宁530029

出　　处：《计算机应用与软件》2023年第9期126-131,136,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61461010);广西自然科学基金重点项目:边缘环境下基于深度压缩的恶意软件检测(2020GXNSFDA 238001);“认知无线电与信息处理”省部共建教育部重点实验室基金项目(CRKL170103,CRKL170104);广西高校云计算与复杂系统重点实验室基金项目(YF16203)。

摘　　要：现实中拍摄的字轮式水表图像的读数区域存在不同的位数和旋转角度的问题,这些问题对识别准确率造成很大的影响。针对这种情况,提出一种基于深度学习的水表读数检测与识别算法。该方法使用改进的圆盘检测算法来对水表圆盘进行检测;采用一种改进的多方向全卷积网络检测出水表的读数区域,同时提出一种图像旋转矫正算法,实现对水表读数区域的矫正与分割;对于读数识别,设计一种轻量级的神经网络,减少模型大小和加速训练过程的同时保持较高的识别精度。实验结果表明,该方法的圆盘检测率从93.97%提高到了96.38%,读数区域检测模型对不同类型的水表读数区域具有较好的检测效果,识别模型的大小从8.77 MB减少到7.32 MB,模型的训练和测试时间短,准确率达到96.44%。In the real world,there are different number of bits and rotation angles in the reading area of the word wheel water meter,which have a great impact on the recognition accuracy.To solve this problem,an algorithm for reading detection and recognition of water meters based on deep learning is proposed.The improved disk detection algorithm was used to detect the water meter disk.An improved multi-direction full convolution network was used to detect the reading area of the water meter,and an image rotation correction algorithm was proposed to correct and segment the reading area of the water meter.For reading recognition,a lightweight neural network was designed to reduce model size and speed up the training process while maintaining high recognition accuracy.The experimental results show that the detection rate of the disk in this method is increased from 93.97%to 96.38%,and the reading area detection model has a good detection effect on different types of water meter reading areas.The size of the recognition model is reduced from 8.77 MB to 7.32 MB,the training and testing time of the model is short,and the accuracy reaches 96.44%.

关 键 词：字轮式水表 深度学习 圆盘检测 图像矫正 轻量级神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]