Development of an artificial intelligence model for image processing of water meters on smartphones, using convolutional neural networks and object detection, using reproduction value monitoring with performance and precision

Abstract

The objective of this research project is to assess the feasibility of developing an artificial intelligence model for processing water meter images using mobile devices. The process of reading a meter is currently carried out manually, susceptible to errors and fraud, therefore, the use of a mechanism that makes automatic the recognition of the information of a meter would make it possible to reduce or mitigate the problems associated to the meter reading, as well as, increasing the productivity of the reader professional.This project will consist of three main stages, the first stage consists of creating a new neural network architecture, based on existing architectures (CR-NET, YOLOv2 and CenterNet), applying optimization and complexity reduction techniques of these architectures, involving the reduction or elimination of convolutional layers of the network, measuring the reduction in processing time versus the decrease in accuracy obtained in each simulation performed. In the second stage, other optimization techniques will be applied, such as the information quantization process (which involves reducing the bit precision of the information used) and the distillation of the neural network (a process that generates a simpler student model from a teacher model that is a complete and more complex one). The last step contemplates the best models that balance the execution time and accuracy and aims to apply specific optimizations for the execution of the neural network on mobile devices. The first two steps involve the development and optimization of the models using the Python language and the Tensorflow framework, and the last step will use the TFLite framework, focused on using artificial intelligence models on mobile devices.The expected result is to obtain an artificial intelligence model that performs the digit recognition of a meter inside a mobile device, instead of having to rely on an external system to carry out this process. Among the secondary results involve the reduction of operational errors and frauds, with the consequent increase in the productivity of the reader professional, making the process of reading the meters, in general, less costly.As an impact, it is expected to be able to penetrate the solution in places where the result needs to be immediate and not post processed. This solution may be deployed for gas and light in the future, justifying the readings growth greater than 100% per year. In addition, the price of reading may be increased from R $ 0.09 to R $ 0.12, being responsible for 25% of Decode's revenue. (AU)