基于不规则目标检测的指针式仪表读数识别方法  

Reading Recognition Method for Pointe Meters Based on Irregula Object Detection

作　　者：潘宇强 姚垚 张林[2] 高俊涛[1] PAN Yuqiang;YAO Yao;ZHANG Lin;GAO Juntao(School of Computer and Information Technology,Northeast Petroleum University;School of Information Management and Artificial Intelligence,Zhejiang University of Finance and Economics)

机构地区：[1]东北石油大学计算机与信息技术学院 [2]浙江财经大学信息管理与人工智能学院

出　　处：《仪表技术与传感器》2024年第6期100-105,共6页Instrument Technique and Sensor

基　　金：东北石油大学特色领域团队专项项目(2022TSTD-03)。

摘　　要：针对目前指针式仪表读数识别方法流程多、累计误差大、对倾斜图像识别效果差的问题,提出一种基于不规则目标检测网络的指针式仪表读数识别方法。首先构建校准网络结构,提取不规则目标顶点坐标,实现对图像自动进行透视变换,强化整体网络对倾斜样本的学习性能;随后利用卷积神经网络直接提取图像特征,实现读数信息的回归任务,减少方法步骤;最后整合模型,使倾斜校准与读数识别任务通过同一个可反向传播的神经网络学习并实现。实验表明,该方法提高了对倾斜仪表图像的读数识别精度,读数流程短、识别效率高。Aiming at the problems of multiple processes,large cumulative errors,and poor recognition performance for tilted imagesin current pointer instrument reading recognition methods,a pointer instrument reading recognition method based on irregular object detection network was proposed.Firstly,a calibration network structure was constructed,irregular target vertex coordinates was extracted,and perspective transformation was automatically performed on the image to enhance the learning performance of the overall network for tilted samples.Subsequently,convolutional neural networks were used to directly extract image features,a-chieving the regression task of reading information and reducing method steps.Finally,the model was aggregated to enable tilt cal-ibration and reading recognition tasks to be learned and implemented together through a backpropagation neural network.The ex-periment shows that the method improves the reading recognition accuracy of inclined instrument images,with a short process and high recognition efficiency.

关 键 词：指针式仪表 不规则目标检测 透视变换 倾斜校准 读数识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]