An Adaptive Robust Pointer Meter Automatic Reading Algorithm Based on CAR-Unet

Significant advancements have been made in automatic reading methods for pointer meters utilizing machine vision. Nonetheless, the accuracy of instrument readings remains susceptible to various factors, including dial contamination, reflections, and tilt distortion. In addition, achieving automatic readings for instruments with different ranges without relying on manual priors continues to pose a significant challenge. Existing methods rarely address issues related to dial distortion and range adaptation. This article proposes a comprehensive pipeline to tackle the aforementioned challenges. First, Yolov8-o is employed to position the dial and numerical indicators, and a coordinate attention based residual Unet (CAR-Unet) semantic segmentation network is introduced to segment the dial and its components. A coordinate attention (CA) mechanism and a residual structure are integrated to enhance the network’s resilience to interferences such as dirt and reflections. Simultaneously, a dial correction method is proposed, utilizing the segmentation mask from CAR-Unet to rectify distortions in the instrument image. Finally, we present a lightweight adaptive range dial number recognition algorithm based on ResNum, which integrates adaptive binarization to facilitate the recognition of multidigit and decimal numbers, enabling the automatic reading of meters with adaptive ranges. Experimental results demonstrate that the proposed pipeline achieves state-of-the-art (SOTA) performance, with a fiducial error of 0.47%. Furthermore, the average processing time for a single image with a resolution of $1920\times 1080$ is only 144 ms.