Abstract: Aiming at the problem of slow speed of intelligent positioning and low accuracy of reading for pointer meters,a method of intelligent reading for pointer meters based on YOLOv5s and U²-Net is proposed.Firstly,the partial batch normalization layer in the trunk and neck of YOLOv5s is used for channel pruning to realize the model lightweight and improve the speed of meter positioning;secondly,in view of shadow and noisy images collected in the field,the image is enhanced,so as to make the model suitable for complex environments;and then,the U²-Net model is adopted to segment the pointer and the scale,to locate the relative position of pointer on the scale,and to improve the accuracy of readings.Then,the U2-Net model is used to segment the pointer and the scale,locate the relative position of the pointer on the scale,and improve the accuracy of the readings;finally,the distance method is used for meter reading.The results of the experiment and reading test show that the size of the lightweight YOLOv5s model is reduced by 56.35%,the positioning time is reduced by 18.18%,the accuracy of the reading recognition test is 93.42%,the average absolute error is 0.032,the average reading error rate is 0.605%,and the average test time per picture is 0.402 s,which meets the industrial inspection requirements.

Key words: pointer instrument, YOLOv5s, channel pruning, U²-Net, distance method