基于YOLOv5的水表字轮读数自动识别方法

Research on Automatic Recognition Method of Water Meter Wheel Indications Based on YOLOv5

基于水表公司内实地采集的一批检定水表数据集,通过数据增强扩增数据集,标注表盘和字轮框数据集并训练模型,实现水表表盘区域和字轮区域定位;标注水表机械字轮的半字符的数据集,在PyTorch框架下搭建YOLOv5算法环境,采用YOLOv5s网络模型进行训练。实验结果表明,该方法的字轮数字识别训练模型的mAP@0.5:0.95值达到0.95,字轮字符识别的整体准确率达到93.85%,与模板匹配方法相比准确率提高了5.58%,其中半字符识别准确率提高了9.15%,有效解决了水表字轮半字符识别错误率高的问题,在水表等仪表读数的自动化改造方面具有一定的应用价值。

A batch of datasets of verified water meters is collected on the spot in water meter company.Datasets is expanded through the image augment method.The frame datasets of dial and word-wheel are marked,and the model is trained to recognize the potion of the dial and word-whel area of the water meter.The half character datasets of the mechanical word-wheel of the water meter are marked.The environment of YOLOv5 algorithm under the PyTorch frame-work is built,and YOLOv5s network model is used for training.The experimental results show that the mAP@0.5:0.95 value of the training model for word wheel number recognition method reaches to 0.95.The overall accuracy of word-wheel character recognition reaches to 93.85%,which is over 5.58%compared with the template matching method.The accuracy of half-character recognition has been improved by 9.15%.It effectively solves the problem of high error rate in semi character recognition of water meter word-wheel,and has certain application value in the automatic transformation of water meter and other meters.