With the rapid development of video image technology and fifth-generation mobile network technology, the automatic verification of mechanical water meters has become an increasingly important topic in smart cities. Although much research has been done on this subject, the efficiency and accuracy of existing water meter pointer reading technology can still be improved. This paper proposes a new water meter pointer reading recognition method based on target-key point detection. Our method consists of a target detection module and a key point detection module. The target detection module uses a modified YOLOv4-Tiny network to detect and classify the areas where the dials and pointers are located in water meter images with distinct characteristics. The key point detection module is used to detect the key point of the pointer image. In this module, the structure of the RFB-Net network is improved to introduce multiple layers of low-level feature information, therefore, it can make full use of the information between multi-scale feature layers for key point detection. In addition to, aiming at the problem of dial rotation, a method of establishing a right-angle coordinate system based on key point is proposed to realize pointer reading. The whole method proposed in this paper is compared to the Hough transform feature matching algorithm and traditional machine learning algorithms through experiments which test the detection and recognition of the water meter dial, pointer and key points. The experiment results show that the missed detection rate of the model in this paper is 1.88% and 1.07% for the dial region and the pointer region, respectively. And the accuracy rate reaches 98.68%, the average processing time per image is 0.37 s. This implies that the water meter inspection task is completed quickly and accurately with strong robustness. Thanks to the lightweight algorithm of our approach, the model can also be fully automated and easily deployed on mobile devices.

随着视频图像技术和第五代移动网络技术的飞速发展，机械式水表的自动检定已成为智慧城市日益重要的课题。虽然在这方面已经做了很多研究，但现有水表指针读数技术的效率和准确性仍有待提高。提出一种基于目标关键点检测的水表指针读数识别新方法。我们的方法由目标检测模块和关键点检测模块组成。目标检测模块使用修改后的YOLOv4-Tiny网络对具有鲜明特征的水表图像中表盘和指针所在的区域进行检测和分类。关键点检测模块用于检测指针图像的关键点。在这个模块中，RFB-Net网络结构改进，引入多层低级特征信息，可以充分利用多尺度特征层之间的信息进行关键点检测。此外，针对表盘旋转问题，提出了一种基于关键点建立直角坐标系的方法来实现指针读数。通过对水表表盘、指针和关键点的检测和识别进行实验，将本文提出的整个方法与霍夫变换特征匹配算法和传统机器学习算法进行了比较。实验结果表明，本文模型对表盘区域和指针区域的漏检率分别为1.88%和1.07%。并且准确率达到98.68%，每张图像的平均处理时间为 0.37 秒。这意味着水表检查任务可以快速准确地完成，鲁棒性强。由于我们方法的轻量级算法，该模型还可以完全自动化并轻松部署在移动设备上。