Desarrollo de un prototipo móvil para la lectura de medidores de agua potable usando técnicas de procesamiento de imágenes.

Abstract

Abstract: In the Ecuadorian context, there are community organizations called Juntas Administradoras de Agua Potable y Saneamiento (JAAPS), which are responsible for managing water and sanitation services in rural parishes and communities. These organizations face challenges when collecting measurement data, as this process is performed manually and is prone to errors that can lead to economic losses and reliability issues. This work presents the development of a mobile Android application prototype for automatic water meter reading using computer vision. Starting from the acquisition of a dataset of 1,749 images of analog meters captured in the field, two YOLOv8 Nano models were trained. The first model was trained to detect the region of interest (ROI) of the counter display, and the second to recognize individual digits, assigning classes from 0 to 9. Both models were exported to TensorFlow Lite for local execution on the device, thus eliminating the dependency on internet connectivity. The application was developed following an offline-first architecture with background synchronization of measurements. Results show that the digit detector achieved a mAP50 of 0.988 on the test set. Field validation, conducted on 200 measurements at a specific JAAP, achieved 99 % accuracy. These results demonstrate the feasibility of automating water meter reading capture in rural contexts.