A Human-Like Reading Method for Pointer Meter Based on Scale Values Recognition and Meter Elements Segmentation

Abstract Due to the simple structure and strong anti-interference ability of pointer meters, they play a crucial role in complex industrial environments. To address the limitations of existing intelligent reading methods, which rely on prior knowledge of scale values, are not adaptable to various types of meters, involve difficulties in meter elements extraction, and are inefficient in reading processes, this paper proposes a human-like reading method for pointer meter based on scale values recognition and meter elements segmentation. Specifically, the YOLODIG algorithm is proposed for the identification of the primary scale values in the scale reading phase, thereby addressing the challenge of scale value a priori, which is applicable to a variety of pointer meters. In the subsequent stage of meter elements extraction, we propose the CL-Multi-U 2 Net multi-class semantic segmentation network to segment the pointer lines and primary scale lines. Our proposed Cross-Scale Aggregation Interaction Module (CSAIM) effectively integrates multi-scale information across layers, increasing the penalty for edge contour segmentation by redesigning the loss function to improve the accuracy of meter elements segmentation. Of particular significance is the proposal of the Human Eye Simulation Reading Method (HESRM) in the reading stage to calculate precise reading values. The HESRM has the advantage of effectively reducing cumulative error and improving reading efficiency and recognition accuracy. The experimental results demonstrate that the YOLODIG algorithm attains an SRM rate of 95.61. The MIOU and F1-score of the CL-Multi-U 2 Net meter elements segmentation network achieve 86.64 and 92.56, respectively, representing improvements of 0.67 and 0.38 compared to the pre-optimization phase. The HESRM proposed in this research attains an average reference error rate of only 0.245%, which is the lowest in comparison with existing methods and is more suitable for practical applications.