Research On Oil Well Detection And Extraction Technology Based On YOLOv3 Model

The development of mineral resources plays an important role in measuring a country's comprehensive strength.With more and more attention paid to the development status of mineral resources and the demand for environmental restoration and management after development,how to monitor the current situation of mineral resources development quickly and efficiently has become an urgent problem to be solved.With the emergence of high-resolution satellite,we can rely on the visual interpretation of remote sensing image to monitor it.However,this method is inefficient,the interpretation time cycle is long and costs manpower,material and financial resources.With the rapid development of the computer field,the computer performance is getting higher and higher,which can carry out the iterative calculation of a large number of remote sensing image data containing massive information.At present,the open Google image resolution can reach decimeters,which contains geometric features such as shape,texture and spatial information of surface objects,which plays an important role in feature extraction of oil well targets.At present,the surface object detection of remote sensing image is mostly for aircraft,ship,automobile,oil tank and so on.There are some public detection data sets for these surface objects,such as UCAS-AOD data set labeled by the pattern recognition Laboratory of National University of science and Technology(including automobile and aircraft samples),and space remote sensing object detection data set labeled by Northwest University of Technology(NWPU Vhr-10,including 10 samples of aircraft,ships,oil tanks,etc.),RSOD dataset annotated by Wuhan University team(including 4 samples of aircraft,playground,etc.),but there is no remote sensing dataset for oil well detection.In this paper,based on Google 0.26 m resolution image data,the oil well target VOC2007 format data set is made,and the target feature extraction of high-resolution remote sensing image is carried out based on the YOLOv3(You Only Look Once)target detection model.Through optimizing the relevant parameters in the model,the sliding slice and edge abandonment methods are proposed to improve the prediction process,and finally the oil well target detection in Google image is realized testing.The main research results are as follows:(1)The improved YOLOv3 model is used to detect and extract oil wells,and finally an accuracy rate of 98.77% is obtained.Compared with the SSD and Retinanet models,the accuracy is higher and the applicability is stronger.(2)The algorithm of sliding slicing and edge rejection is proposed,which solves the problem of repeated detection of large-scale images and improves the accuracy of model detection.(3)The detection results are displayed in the form of a vector shapefile,and the x,y coordinates and confidence information of each predicted point are written into the attribute table in the form of attributes,which improves the practical application significance of the model.