Automatic meter reading via simulated water meter wheel rotation data generation

With the rapid development of smart cities, automated meter reading (AMR) technology, as a crucial part of the infrastructure, has garnered widespread attention. This paper addresses the issues of data insufficiency and imbalanced sample distribution in AMR technology by constructing a simulated water meter digit generator that automatically generates water meter data images and their corresponding label files. Furthermore, to tackle the challenge of semi-character recognition, the model's learning capability is enhanced by increasing the semi-character data samples and introducing auxiliary features. Finally, a counting area extraction method that combines object detection algorithms with perspective transformation techniques is introduced into the existing AMR framework, effectively resolving problems related to image skew correction and perspective distortion. Experimental results demonstrate that, compared to the original dataset, training with the enhanced dataset proposed in this paper significantly improves the model's performance in recognizing water meter digits, particularly showing advantages in semi-character recognition, thereby providing new insights and solutions for the development of AMR technology.