Full-automatic Reading Recognition for Pointer Meters Based on Improved Rotating Object Detection Model
Aiming at the difficulty of accurate positioning of tick marks and pointers in pointer meter images,as well as the problem of false detection and missed detection under complex environment,a full-automatic reading recognition method for pointer meters based on an improved rotating object detection model was proposed.Firstly,based on the YOLOv5s network,an efficient channel and spatial attention fusion module were designed to improve the ability to extract reading-related features on dial image.Secondly,E-CIoU Loss was designed to optimize the loss function and enhance the ability to regress the pointer bounding box.At the same time,circular smooth label(CSL)was introduced to adapt to the task of rotating object detection.Then,the improved probabilistic Hough transform was used to achieve precise pointer repositioning.Finally,the reading recognition result was calcu-lated by using the relative position relation of pointer and tick marks in polar coordinate plane.Experimental results indicate that the proposed method effectively improve the detection accuracy of the reading-related features on dial image with the mAP value of 96.8%when compared with the benchmark model,and the average relative error of the final reading recognition reaches 0.52%,which can meet the practical application requirements.
rotating object detectionattention mechanismsoptimization loss functionimproved probabilistic Hough transform
NETL
NSTL
万方数据
