A SOLUTION FOR AUTOMATED WATER METER READING FROM IMAGES BY APPLYING DEEP LEARNING

Abstract

In this paper, we propose an automated meter reading (AMR) method applied to water meters by applying deep learning. We design a two-stage method using Rotational Region Convolutional Neural Networks (R2CNN). The first stage uses a R2CNN network to detect the digit region; the second stage applies another R2CNN network to recognize digits on an image that has only the alphanumeric region. The digits after identification are processed and sorted to obtain the counter meter. In ARM studies, most datasets are not available to the research community because the images belong to service companies. Therefore, in this study, we created a new dataset for the proposed method using it for training and testing. The result is a process with deep learning models that determine water meter readings from images of the watch face with high accuracy, and this process has been integrated into Citywork software to initially help developers. The manager audits the accuracy of the readings recorded by the employee manually to see if they match the readings in the images.