基于深度学习的指针式机械水表读数识别算法  

Pointer Water Meter Reading Recognition Algorithm Based on Deep Learning

作　　者：薛振豪 许书君[2] 周哲帆 王敏[1] 文向 喻珺岩 郭玉彬[1,3] 李西明 XUE Zhenhao;XU Shujun;ZHOU Zhefan;WANG Min;WEN Xiang;YU Junyan;GUO Yubin;LI Ximing(Mathematics and Informatics College,South China Agricultural University,Guangzhou 510642,China;Digital Security Academy,Shandong College of Electronic Technology,Ji'nan 250200,China;Guangzhou Smart Agriculture Key Laboratory,Guangzhou 510642,China)

机构地区：[1]华南农业大学数学与信息学院,广东广州510642 [2]山东电子职业技术学院数字安全学院,山东济南250200 [3]广州市智慧农业重点实验室,广东广州510642

出　　处：《软件导刊》2025年第2期163-171,共9页Software Guide

基　　金：国家自然科学基金项目(61872152);广州市科技计划项目(201902010081)。

摘　　要：指针式机械水表主要依靠人工进行抄表和识别,存在耗时长、人工成本高、识别错误率高等缺点。近年来随着深度学习技术的发展,研究人员将其应用于水表读数识别方面。设计一套基于深度神经网络的指针式机械水表读数识别算法(PWMR-DL),可准确地识别指针式机械水表的读数,并构建指针式机械水表数据集用于算法训练和测试。针对子表盘的检测和矫正,引入MaskRCNN模型实现表盘定位与分割,并设计了高效的矫正策略对各个子表盘进行旋转校正,以提升指针式机械水表图像在不同旋转角度下识别的鲁棒性,减少误差。在子表盘读数识别阶段,引入CA注意力机制改进EfficientNet模型,以提升读数识别的准确率,并通过增加分类维度到20类,细化了指针位置处于数字间隙时的判断精度。同时,结合子表盘读数序列相关性校正逻辑设计读数生成方法,有效减少了读数错误。实验结果表明,PWMR-DL算法在子表盘读数识别方面,与改进前的EfficientNet模型相比精度提升了约2.4%,而且经过优化的模型仅增加了少量参数,维持了其轻量级的特性。在低分辨率图像下,PWMR-DL算法的整体识别精度可达到96.8%。Traditional pointer-style mechanical water meters predominantly rely on manual reading and recognition processes,which are often time-consuming,incurs high labor costs,and are prone to high error rates.With the evolution of deep learning technologies,researchers have been applying these advancements in the field of water meter reading recognition.In this study,we propose a deep neural network-based algorithm for the recognition of readings from pointer-style mechanical water meters,referred to as the PWMR-DL algorithm.A specialized dataset for pointer-style mechanical water meters was constructed for the training and testing of the algorithm.To detect and correct for sub-meter dials,the Mask R-CNN model was employed to locate and segment the dials,coupled with an efficient correction strategy to rotationally adjust individual sub-dials,thereby enhancing the robustness of recognition across various rotational angles and reducing errors.During the sub-dial reading recognition phase,the CA(Channel Attention)mechanism was introduced to refine the EfficientNet model,which significantly improved reading accuracy.By increasing the classification dimension to 20 classes,the algorithm refines the precision of judgments when the dial pointer is situated between numerals.Furthermore,by incorporating a correction logic related to the sequence of sub-dial readings,an effective method for generating readings was designed,substantially reducing errors.Experimental results demonstrate that the PWMR-DL algorithm achieves a 2.4%increase in precision for sub-dial reading recognition compared to the pre-improved EfficientNet model,while only incorporating a small number of additional parameters,thereby preserving the model's lightweight characteristic.Notably,the PWMR-DL algorithm attained an overall recognition accuracy of 96.8%even under low-resolution imaging conditions.

关 键 词：计算机视觉 EfficientNet 指针式水表 读数识别 CA注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]