Research On Oil Field Instruments Reading Recognition Algotithm Based On Deep Learning

In recent years,the biggest change to oilfield production site inspection is to use intelligent inspection robot instead of traditional manual inspection.The reading recognition of instruments is one of the main inspection tasks of inspection robot.Because of the complex environment of oil field,the task becomes important and difficult.Therefore,it is very necessary to study the oil field instrument reading recognition algorithm based on deep learning in this paper.There are still some problems in the existing instrument reading recognition algorithm,mainly including the following aspects.(1)Firstly,the existing traditional reading recognition methods for pointer instruments have the problems of cumbersome steps and unstable performance.And the current discrete reading recognition method based on deep learning classification network has great limitations.(2)Secondly,the existing traditional reading recognition algorithm for digital instruments is not only easy to be disturbed,but also the recognition effect is unstable.Some researches also use deep neural network,but it needs a large number of data sets.Therefore,this paper proposes the research on oil field instrument reading recognition algorithm based on deep learning.The specific research contents are as follows.(1)Based on SSD algorithm completes the preliminary positioning and classification of instruments.The algorithm can simultaneously realize the positioning and classification of the instrument with a small amount of calculation,and make preliminary preparation for the reading recognition of the two kinds of instruments.(2)The super-resolution reconstruction of fuzzy instrument image collected by inspection robot is studied based on SRGAN algorithm.After the instrument image is collected,the instrument image may be unclear due to factors such as camera shaking or weather when the inspection robot is shooting,which will affect the subsequent reading recognition.Therefore,SRGAN network is proposed to carry out super-resolution reconstruction of the fuzzy image collected.(3)The reading recognition of pointer instrument in oil field is studied.Aiming at the existing problems,a kind of classification coarse positioning combined with regression fine reading method is proposed to realize the reading recognition function of pointer instrument.This method not only simplifies the instrument reading recognition step,but also effectively solves the depth of the current classification network can only identify discrete integers reading problems.(4)The reading recognition of oilfield digital instrument is studied.Aiming at the problems of existing algorithms,a digital instrument reading recognition method based on deep transfer learning is proposed.The idea of transfer learning is introduced into Resnet18 network,so that the network can carry out adaptive learning on source domain in advance,and transfer the feature information to the segmented character image to complete the training of the model.Then deformable convolution is added to the model to further enhance its robustness.In order to verify the effectiveness of the algorithms proposed in this paper,experiments are carried out on them respectively,and the experimental results are as follows.In the positioning and classification of the two instruments based on SSD algorithm,the mAP value of 0.98 is achieved.In the super-resolution reconstruction experiment of fuzzy instrument image based on SRGAN algorithm,the PSNR value of 31.415 and SSIM value of 0.993 are obtained.In the reading recognition algorithm of pointer instrument.In the first stage,the accuracy of the rough classification model reached 97.38%,and in the second stage,the accuracy of the regression models corresponding to the four regions are 91.72%,94.06%,91.54%and 92.84%respectively.In the digital instrument reading recognition algorithm based on the idea of deep transfer learning.The accuracy of resnet18 network with transfer learning is 98.64%,and the accuracy of the model with deformable convolution is improved to 99.53%.To sum up,the method proposed in this paper can meet the needs of patrol inspection in actual oil fields.