基于对抗学习的指针仪表自适应读数识别算法

Adaptive reading recognition algorithm of pointer instrument based on adversary learning

针对指针仪表采用人工读数方式存在的成本较高、读数不准确、时效性较差的问题,提出一种基于对抗学习的指针仪表位姿自适应读数识别算法。该算法通过目标检测模型识别图像中的指针仪表的位置和姿态,将指针仪表进行校准后并利用数字图像处理技术进行读数识别。为了提升目标检测模型对位姿不同的指针仪表图像的识别效果,本文提出了基于对抗学习的数据增强技术,通过优化搜索模型识别不准的图像旋转角度、平移距离以及缩放比例构造训练数据,提高目标检测模型在指针仪表位姿发生变化时的准确率。以工矿企业中常用的SF6压力仪表为实验对象,实验结果表明读数识别的误差在1%以内,证明了所提算法的有效性。

To address the problems of high cost,inaccuracy,and low efficiency in manual reading of pointer instruments,a pose-invariant adaptive reading recognition algorithm of pointer instrument based on adversarial learning is proposed.This algorithm utilizes an object detection model to identify the position and attitude of the pointer instrument in the image,calibrates the pointer instrument and uses digital image processing technology for reading recognition.In order to improve the recognition effectiveness of the object detection model on pointer instrument images with different poses,a data augmentation technology based on adversarial learning is proposed,which constructs training data by optimizing rotation angles,translation distances,and scaled ratios of images that lead to inaccurate recognition,to improve the accuracy of the object detection model when the pointer instrument′s pose changes.The research focuses on the SF6 pressure instrument commonly used inindustrial and mining enterprises,and experimental results show that the error of reading recognition is within 1%,which proves the effectiveness of the proposed algorithm.