Detection and Recognition Methods for Pointer Meters Under Interference Conditions With Improved Ellipse Fitting Methods

The advancement of intelligent technology has heightened the demand for efficient data acquisition. Pointer meters are easily affected by environmental factors and tilted shooting angles, causing existing detection and recognition methods to degrade in industrial applications. To address this, we propose a highly reliable detection and recognition method for pointer meters under interference conditions. We designed the CSP-Rep-Att-C2F-Yolox (CRAC-Yolox) object detection module using reparameterization tricks and attention mechanisms to enhance detection speed and accuracy while mitigating environmental interference. Furthermore, we proposed the polarized self-attention and feature fusion Deeplabv3+ (PAF-Deeplabv3+) to segment key components of the meters, leveraging the polarized self-attention module (PSAM) and feature fusion module (FFM) to improve robustness and accuracy. The improved ellipse fitting method uses segmentation results to correct the center offset overlooked by traditional methods during perspective distortion correction. Finally, the reading value is calculated based on the angle between the pointer and the two nearest main scales. The experimental results show that CRAC-Yolox achieves 84.8% mAP50:95 detection accuracy, and PAF-Deeplabv3+ achieves 85.57% mIoU, outperforming other state-of-the-art (SOTA) methods. Our pointer meter detection and recognition (PMDR) method achieves a reading success rate of over 90%, with an average relative error of 1.498% and an average reference error of 0.439%, meeting industrial application accuracy requirements. It also maintains high success rates despite various environmental interferences. In addition, we propose the PMDR dataset for training and testing all intermediate modules of the pointer meter reading recognition method.