Pengenalan Karakter Optis untuk Pencatatan Meter Air dengan Long Short Term Memory Recurrent Neural Network

Abstract

Clean water service providers in Indonesia are still recording water meters as water usage data with manual recording by record collector. Alternative solutions for recording water meters from previous research use the Internet of Things (IoT) or image recognition that is processed on a server. The solutions rely on the Internet which is unsuitable with Indonesia’s condition. This study proposes a water meter reading system that can work on mobile devices without using the Internet. The system works by utilizing optical character recognition (OCR) using the Long Short Term Memory Recurrent Neural Network (LSTM-RNN) method. LSTM-RNN is a classification method in artificial neural network which has feedback. The results show that the water meter reading system could work without using an Internet connection. The average time it takes to perform the reading process is 2285ms even on Android device with low specification. The overall reading accuracy is 86%. Single value reading accuracy, when the digit meter displays only 1 number, is 97%, while the accuracy of double value reading, when the digit meter displays 2 numbers, is 18%.