Research on the Pointer Meter Reading Method Based on Adaptive Pooling and Hierarchical Fusion

Abstract

A pointer meter reading method based on adaptive pooling and hierarchical fusion is proposed, aiming at the current pointer-type meter reading methods’ poor recognition ability and low accuracy. Firstly, to tackle the limited recognition capability of dial areas in complex settings, we proposed a pooling cascade and a lightweight instrument detection algorithm based on DA-YOLOv8n. The approach aims to enhance the feature detection ability of challenging-to-identify dials by supplementing feature information from various receptive fields for small object detection and boosting detection speed. Secondly, the AF-U2Net hierarchical fusion algorithm is proposed for the segmentation of scale lines and pointer, which extracts the fine-grained information of edge pixels through multilevel layering and multiscale feature splicing to alleviate the problem of low segmentation accuracy effectively. Finally, precise meter readings at the pixel level are attained through refinement operations, polar coordinate transformations, and projection operations. Experimental results indicated that DA-YOLOv8n improved compared to the original model, with an increase of 1.6% in [email protected] and 1.2% in [email protected]. AF-U2Net’s MPA improved by 7.42%, and MIoU enhanced by 1.30%. The average quoted error of this reading method is 0.47%, which is promising in the field of pointer meter reading.

Keywords

• Deep learning
• indication recognition
• image segmentation
• object detection
• pointer-type meter