Research On The Recognition Method Of Pointer Meter Readings Based On Deep Learning

Pointer instruments are measuring instruments widely used in industrial fields,such as pressure gauges,temperature gauges,water level gauges,etc.However,due to environmental,technical and other limitations,most industrial analog meters still take the way of manual reading.With the development of computer vision technology and image processing technology,the automatic reading of the value of the pointer meter has become a research hotspot.However,the current automatic reading algorithm does not have high reading accuracy in natural scenarios,and the application effect is not good in industrial scenarios(substations,gas stations,etc.).Therefore,based on deep learning,this paper studies the automatic reading method of pointer meters,and strives to study a method that can perform high-precision automatic meter reading in natural scenes.The specific research contents are as follows:First,the research on the detection model of the pointer instrument panel.First,this paper uses the improved algorithm based on the target detection algorithm Yolov4 to locate the target meter and enlarge the target table.Secondly,network structures such as atrous convolution,Densenet,and SAM are introduced to improve the accuracy of the model under natural conditions.After experimental verification,the dial detection model proposed in this paper can achieve 93.27% accuracy on the dial data set collected in natural scenes,which is higher than the original Yolov4 model.Second,research on the identification method of pointer meter readings.First,the pointer region is extracted using the Unet-based semantic segmentation model,and through this pointer region,the pointer angle is fitted.Second,the character detection algorithm CRAFT and the text recognition algorithm E2E-MLT are fused and applied to recognize scale text and units on tables.At this stage,by identifying the position of the dial scale text,the center of the circle is located for subsequent polarization.After that,the scale area of the table is converted to the polar coordinate system,and a lightweight convolutional neural network is designed to locate the main tick mark corresponding to each scale.Finally,an algorithm is designed to combine the above information to calculate the readings.Third,model application validation.First,in order to train and verify the deep learning model proposed in this paper,this paper collects images of pointer meters from real industrial scenarios such as gas stations,oil fields,and power plants,and classifies them to construct a dataset.Afterwards,each deep learning model is trained and validated,and images are read to compare with previous automatic meter reading algorithms.Finally,the automatic reading method of pointer meter based on deep learning proposed in this paper is verified in real scenarios.The method in this paper is deployed on a mobile inspection robot,and the reading test is carried out in a largescale industrial scene,and the effectiveness of the method is verified by comparing the manual reading.