Research On Pointer Instrument Recognition Method Based On Deep Learning

The digital transformation and intelligent development of the oilfield are the guarantee of efficient management,energy conservation,production increase and safe operation of the oilfield.As a task of intelligent oil fields,unmanned inspection mainly utilizes equipment such as drones and robots to replace human labor in daily inspection work through visual management,and it ensures the safety production of oil and gas fields.At present,pointer instruments are commonly used for parameter monitoring in oil fields,and the pointer instruments are recorded by manual inspection.But there are some issues such as human reading errors,inability to record and upload data in real-time,or existing blind spots in remote areas.Using the machine vision technology in the automatic recognition of pointer instrument can achieve the unmanned intelligent inspection technology of oil fields.Researches have shown that the automatic recognition method of pointer instrument readings is based on digital image processing,which is achieved by analyzing,extracting features,and calculating readings from instrument images collected by industrial cameras.However,the common methods do not consider the impact of the on-site environment on image quality,and it does not have good robustness to against noisy and foggy factors.Therefore,in order to solve the problem of automatic recognition and reading methods for pointer instruments,this paper proposes to use the method based on deep learning to research image preprocessing methods for pointer instruments,the methods for recognizing,segmenting,and reading.In this study,Because the noise during image acquisition or transmission and the fog in the environment have a certain impact on the accuracy of recognition,this paper proposes an image denoising algorithm based on wavelet function and an image defogging algorithm based on cycle generation countermeasure network to preprocess the image.By using this method to preprocess the image,the clarity of the image is enhanced,ensuring the effectiveness of instrument image recognition.And then using YOLOv5s which can test faster and deploy easier detects the dial and pointer information of clear instrument panels.Next,use the trained DeepLabv3+network to segment the image,extract pointer regions,and achieve feature extraction of the instrument pointers.Finally,the PCA refinement method is used to refine the pointer area,and using the angle method calculates the angle between the refined pointer and the zero scale line to obtain an accurate scale.Based on the experimental results,it shows that the average absolute error of the included angle of the pointer obtained in this paper is 0.33°,the average absolute error of the indicator recognition is 0.1 ℃,and the average relative error is 0.12%,which proves that this method is more accurate and more applicable in the automatic recognition of instrument readings in the environment of noise and fog.