基于BSRGAN的集气站仪表读数识别  

BSRGAN-Based Meter Reading Recognition at Gas Gathering Stations

作　　者：刘金海 祝忠钲 范旭[1] 邱瑞宸 葛笑 韦涛[1] 常雅晴 肖华旗 LIU Jinhai;ZHU Zhongzheng;FAN Xu;QIU Ruichen;GE Xiao;WEI Tao;CHANG Yaqing;XIAO Huaqi(Oil Production Service Branch,CNOOC Energy Development Co.,Ltd.,Tianjin 300452,China;School of Automation,Southeast University,Nanjing 210096,China)

机构地区：[1]中海油能源发展股份有限公司采油服务分公司,天津300452 [2]东南大学自动化学院,江苏南京210096

出　　处：《自动化仪表》2023年第8期7-14,共8页Process Automation Instrumentation

基　　金：国家自然科学基金资助项目(61973083)。

摘　　要：为解决集气站的巡检机器人采集到的仪表图像分辨率低、常伴有噪点和仪表图像畸变、不便于读数算法直接应用等问题,提出了一种基于盲超分辨率生成对抗网络(BSRGAN)的仪表读数算法。首先,利用BSRGAN提高输入仪表图像的分辨率,较传统方法能够显著增强仪表图像的视觉感知水平。自然图像质量评估(NIQE)在测试集中表现较好,结构相似性指数(SSIM)值均大于0.85。利用尺度不变特征变换(SIFT)算法对重建图像进行特征提取。特征点数量在重建后显著增加,有助于表盘图像的校正。然后,采用连通域划分和指针细化等形态学操作提取指针中心线。最后,利用角度法识别表盘读数。试验读数结果与真实值的平均相对误差为0.625%。试验结果表明,所提算法可行、有效、精准,能够提高仪表图像清晰度和视觉感知,适用于集气站及其他复杂环境的仪表读数。In order to solve the problems of low resolution,often accompanied by noise and instrument image distortion,and not convenient for the direct application of reading algorithms of the instrument image collested by inspection robot of the gas gathering stations,the instrument reading algorithm based on blind super-resolution generative adversarial network(BSRGAN)is proposed.Firstly,the resolution of the input meter image is improved using the BSRGAN,which can significantly enhance the visual perception level of the meter image compared with the traditional method.The natural image quality evaluator(NIQE)performs better in the test set,and the structural similarity index measure(SSIM)values are all greater than 0.85.The reconstructed images are feature extracted using scale invariant feature transform(SIFT)algorithm.The number of feature points is significantly increased after reconstruction,which helps in the calibration of the meter dial images.Then,morphological operations such as connectivity domain division and pointer refinement are used to extract the pointer centerline.Finally,the dial readings are identified using the angular method.The average relative error between the test reading results and the true value is 0.625%.The experimental results show that the proposed algorithms is feasible,effective and accurate,can improve the instrument image clarity and visual perception,and is suitable for instrument readings in gas gathering stations and other complex environments.

关 键 词：集气站 巡检机器人 指针式仪表 超分辨率重建 深度学习 图像校正 仪表识别 

分 类 号：TH701[机械工程—仪器科学与技术]