摘要
为推动水务行业智能化发展,实现对机械水表的自动化识别,本研究提出了一种同时适用于指针式和转轮式机械水表的读数识别算法。首先,利用市面上常见的水表样本构建了一个水表数据集,然后基于此水表数据集对SSD目标检测算法进行了改进。实验结果显示,经过改进的SSD目标检测算法在模型体积上实现了70%的压缩,同时mAP指标达到了99.31%,提升了1.4%,此举在保证性能的前提下有效减小了模型的体积。此外,还针对网络的识别目标进行了巧妙的设计,将转轮数字和指针的识别结果分为正常状态和进位状态。采用相同的后处理程序,得到了最终的读数结果,这一策略巧妙地解决了两种不同类型的机械水表读数需要两种独立算法分别识别的问题,为水表的自动化识别提供了创新性的思路和实用价值。
In order to promote the intelligent development of the water industry and achieve automatic recognition of mechanical water meters,this study proposes a reading recognition algorithm suitable for both pointer-type and rotary-type mechanical water meters.Firstly,a water meter dataset was constructed based on common water meter samples in the market,and extensive data augmentation techniques were applied to expand the dataset.Subsequently,the SSD object detection algorithm was deeply improved.Experimental results show that the improved SSD object detection algorithm achieved a 70%reduction in model size while achieving a mAP(mean average precision)score of 99.31%,an improvement of 1.4%.This effectively reduces the model size while ensuring performance.In addition,a clever design was applied to recognize the targets of the network,dividing the recognition results of rotary digits and pointers into normal and carry states.Using the same post-processing program,the final reading results were obtained.This strategy ingeniously solves the previous problem of requiring two independent algorithms for recognition,providing innovative ideas and practical value for the automatic recognition of water meters.
作者
丁晓嵘
耿艳兵
DING Xiaorong;GENG Yanbing(Beijing Smart Water Development Research Institute,Beijing 100036,China;Computer Science and Technologyof The North University of China,Taiyuan 030051,China)
出处
《自动化与仪器仪表》
2024年第9期122-126,130,共6页
Automation & Instrumentation
关键词
水表读数识别
目标检测算法
轻量化网络
水表进位
water meter reading recognition
object detection algorithm
lightweight network
water meter carry
