Water Meter Reading Recognition Based on Deep Learning
Yuhao Shen, Wang Xiang-qian, Minghong Yin · 2023
The widespread and large-scale distribution of water meter users in China poses a significant challenge to the traditional manual meter reading method in most regions of the country. Rapid and accurate meter reading for a large number of water meters distributed in different areas has become an urgent problem for various water supply companies. With the rapid development of deep learning, computer vision and other fields, more and more industrial and commercial projects are starting to utilize machine learning-based computer vision technology to process water meter readings, which previously required manual identification by human eyes.This paper mainly studies the problems of text area extraction and reading recognition of the dial-type water meter in real scenarios. Based on deep learning methods, DBnet semantic segmentation algorithm is first used for text detection, followed by CRNN convolutional recurrent neural network for text recognition. Regarding the difficulty of recognizing half characters, this paper proposes an agent character scheme by readjusting the format of the training set, which significantly improves the accuracy of half character recognition. The model is trained and tested on a dataset of real-scenario dial-type water meter images, and the test results show that the algorithm network still has good accuracy and fast recognition speed in complex scenarios such as low light, reflection, partial occlusion, angle tilt, and half-characters.