Thailand Water Meter Reading Using Convolutional Neural Networks from Smartphone Imagery

EasyChair Preprint 11882

Abstract

Water services are supplied by governmental agencies or private firms in many nations, with water delivered to houses via a network of pipes known as waterworks. The households typically consume water over a period of a month, after which the authority in charge evaluates the total amount of water consumed by each household and provides matching invoices. Thus, every household must install a water meter to ensure precise water consumption measurement. In Thailand, the procedure of collecting and reporting readings from these water meters is currently done manually by staff, which can be time-consuming and error-prone. The research's fundamental purpose is to develop a detection model capable of accurately identifying numbers on water meters within images captured using mobile devices. Convolutional Neural Networks (CNNs) were employed as the primary methodology for constructing the desired predictive models. This research is specifically dedicated to the task of reading values from water meters used in Thailand, encompassing five distinct models of water meters. The accuracy of these models is categorized into two components: one for precisely locating a set of numbers on water meters with a precision rate of 99.7%, and the other for accurately reading these numbers with a precision rate of 96.60%.

Keyphrases: Automated water meter reading, Water meter reader, deep learning, machine learning

@booklet{EasyChair:11882,
  author    = {Nattanon Saetan and Kwankamon Dittakan},
  title     = {Thailand Water Meter Reading Using Convolutional Neural Networks from Smartphone Imagery},
  howpublished = {EasyChair Preprint 11882},
  year      = {EasyChair, 2024}}