Water meter reading recognition method based on character attention mechanism

Abstract

With the rapid advancement of computer vision technology, traditional manual methods of reading meters are increasingly being replaced by automated water meter reading technologies based on image recognition. This technology can precisely locate and recognize the readings on captured images of water meter dials, laying a solid technical foundation for the implementation of remote automatic meter reading systems. However, in practical applications, the recognition of water meter readings still faces challenges due to interference from factors such as shooting angles and changes in environmental lighting. To address these challenges, this paper proposes an innovative method based on deep learning. Firstly, the ResNet-based Feature Pyramid Network (FPN) is used to detect the reading area of the water meter to ensure the accuracy of the detection. For the problem of digit character detection, the character detection attention mechanism is introduced to improve the performance of digit detection and reduce the interference of background noise while ensuring high accuracy. For numerical character recognition, the improved LeNet-5 network can better identify water meter readings in natural scenes. Additionally, the integration of a global average pooling layer within the network effectively alleviates the issue of overfitting. To verify the effectiveness of our method, we conducted experiments on the CCF real-world water meter reading automatic identification dataset. The experimental results show that by scaling the water meter reading area and introducing the character attention mechanism to assist in numerical character detection, the recognition accuracy of individual digits improved by 8.8% and 5.5%, respectively, and the overall recognition accuracy of the final water meter reading also increased by 7.0% and 2.2%. These significant improvements demonstrate the superiority and effectiveness of our method in practical applications.