Improved YOLO-v5s Model for Substation Pointer Meter Reading

To promote intelligent control and monitoring of substation instruments and address challenges in deploying universal detection and recognition models on mobile platforms for automatic readings of analog meters in complex backgrounds, a substation-specific meter detection and recognition network was constructed based on YOLOv5s. This work proposes a novel substation analog meter recognition algorithm, PM-YOLOv5s. By leveraging FasterNet as the feature extraction network, the model's computational load was reduced while enhancing feature representation capabilities. To balance operator representational power with hardware computation costs and improve model processing speed, a distributed shift convolution (DSConv) was introduced to optimize the CSP2_3 module. Additionally, the Focal EIOU loss was employed to balance positive and negative samples, thus improving the model's accuracy in recognizing small meter targets. Experimental results demonstrate that the modified YOLOv5s achieves an average precision of 98.51%, with parameter count reduced by 30% and recognition speed increased from 27.817 f/s to 56.968 f/s. These results indicate that the model's real-time performance and accuracy meet the engineering requirements for intelligent data acquisition, detection, and recognition in substations.