Low-Cost Remote Reading of Analog Electric Meters Using Computer Vision and Neural Networks

Abstract

This article presents the design and implementation of an automated remote reading system for analog electricity meters, based on computer vision and artificial neural networks. The ESP32-CAM module is used as the capture unit, evaluating three optical configurations of the OV2640 sensor (66°, 120°, and 120° adjustable) to identify the most suitable in terms of sharpness and coverage. The obtained images are processed using binarization, segmentation, and normalization techniques and input into a neural network trained with the MNIST dataset for automatic digit recognition. The results show that the 120° adjustable variant at a distance of 5 cm offers greater capture accuracy. The neural model achieved an accuracy of 96.18% when evaluated with external images, demonstrating its generalization capabilities in real-world conditions. This proposal represents a low-cost, replicable, and technically viable solution for environments where analog meters are still used, contributing to energy automation in contexts with limited infrastructure.