【作者】 杨帆；

【导师】 金连文；

【作者基本信息】 华南理工大学 ， 信号与信息处理， 2019， 硕士

【摘要】 水资源是关系到国计民生的基础资源,我国存在严重的水资源短缺、水污染加剧和水土流失等问题。随着水资源的日益消耗,对个人用户及企业用户的水表读数进行准确记录具有重要意义;另一方面,随着人口的增长和社会经济的发展,目前的水表用户数量庞大、分布范围广泛,人工抄表的难度越来越大,已经不再适应于当前时代发展的需求,智能抄表系统已成为水务企业目前及未来的发展重点。随着现代社会计算机技术、通信技术的进步,自动化的智能抄表已经逐渐成为现实。本文针对字轮式的水表,提出采用计算机视觉以及图像处理等技术对水表图像进行数字区域的检测以及读数的识别。采用图像识别的方式记录水表读数主要是依托于近几年计算机视觉特别是深度学习技术的高速发展使得复杂环境下的图像检测以及识别已经成为可能,并且将水表读数以图像的方式进行记录可以有效避免作假事件的发生。本文的主要工作包括:1)提出了一个适用于水表读数识别的全卷积序列识别网络,构建并发布了一个用于训练水表读数识别算法的数据集。该数据集包括5000张困难样本和1000张简单样本,困难样本中存在较大的光照、模糊、水滴及灰尘干扰。这两部分的样本都进行了读数标注,如“0,1,5,6,0”。2)针对单个水表数字存在“中间状态”的问题,根据实际应用中对水表读数后处理的分析,提出了一种“增强损失函数”。该损失函数指导模型将处于“中间状态”的数字识别为相对应的“较低状态”的数字,从而减少读数识别的错误,能有效提升识别的准确率。3)基于图像检测、文本识别算法以及云计算技术构建了一个智能水表云服务系统。将水表的图像及读数识别结果保存在云端服务器,同时提供查询和纠错等功能,以达到提高效率和降低人工成本的目的。4)设计并实现了水表读数识别在智能移动终端上的应用演示系统。在服务器上训练了8比特轻量级的水表数字区域检测模型以及水表读数识别模型,并将这两个模型部署到智能移动终端上,开发了一个安卓手机上的演示应用。更多还原

【Abstract】 Water resource is the basic resource related to the national economy and people’s livelihood.With the increasing consumption of water resource,it is of great significance to accurately record the water meter reading of individual and enterprise users.On the other hand,with the growth of population and the development of social economy,water meter users are widely distributed.Therefore,manual water meter reading is becoming more and more difficult,and has no longer adapted to the needs of the current era.The intelligent water meter reading system has become the current and future development focus of water enterprises.With the process of computer and communication technology in modern society,automatic intelligent water meter reading has gradually become a reality.In this paper,we propose to detect the digital area and recognition the reading of the water meter image by using computer vision and image processing methods.And the capture of water meter images can be used to prevent cheating on water consumption.The main work of this paper includes:1)A “fully convolutional sequence recognition network” for water meter reading is proposed and a data set for training is presented.The data set includes 5,000 difficult samples and 1,000 easy samples.Within the difficult samples,there is a wide range of variation caused by illumination,refraction and occlusion.Both the difficult and easy samples are labeled with sequential characters such as ‘‘0,1,5,0,6’’.2)To deal with the problem of “intermediate state” for a single water meter number,an “augmented loss function” is proposed.This loss function instructs the model to learn the number at "lower state" rather than "intermediate state",thereby reducing errors in recognition and effectively improving the accuracy of recognition.3)An intelligent water meter cloud service system is constructed based on image detection and text recognition algorithm.By saving the water meter images and recognition results in the server,we provide useful functions such as image query,recognition result query and correction,so as to reduce manual labor cost.4)An application demonstration system for water meter reading on smart mobile devices was designed and implemented.The 8-bit lightweight quantized digital area detection model and reading recognition model were trained on the gpu machine and deployed on an android device as a demonstration application.更多还原