Real-Time Water Meter Reading Based on TinyML

This paper proposes a system that automatically reads the value of the water meter using a tiny machine learning model (tinyML) running directly on the ESP32-CAM kit. A convolutional neural network (CNN)-based tinyML model is recommended to recognize meter digits before estimating a value. The digit recognition machine learning (ML) models were trained using three datasets generated from this study, including a grayscale image set, a contrast-enhanced image set using the HE algorithm, and a binary image set using the threshold determined by the adaptive threshold (AT) algorithm to find the matching set of images. The experimental results show that the proposed classification model with the input gray image and contrast-enhanced image gives the best accuracy of 98.3%, and the estimated speed is approximately 3 times per second. This accuracy is approximate compared to the previous study; however, the image data processing solution in this study provides roughly 10 times faster estimation time. Furthermore, the study shows that gray images should be used directly for the digit classification problem instead of being contrast-enhanced or converted to binary Image.