An Interactive Rock Slice Image Analysis Method Based On Machine Learning

The information of mineral composition and microscopic pore structure contained in the rock thin section is of great significance to the analysis and prediction of oil and gas reservoirs.Therefore,the identification of rock thin section is one of the important tasks in petroleum geology and exploration and development of oil and gas fields,affecting the productivity of oil and gas wells and the ultimate recovery factor.It is an important guarantee for the efficient exploration and development of oil and gas reservoirs.The traditional identification method is to place the rock thin section under the polarizing microscope.Through manual observation and identification,the workload is large and vulnerable to external factors.The identification result is often inaccurate,and there is even a certain misjudgment rate.Ultimately affect the identification of oil and gas reserves.Through machine learning and interactive image analysis methods to identify complex rock thin section images,the time cost and error rate of manual identification are reduced,the recognition efficiency is greatly improved,and it can be accurate to identify each pixel in the image.It provides an important foundation for further oilfield reservoir development.First of all,the experiment will select suitable rock thin section samples that are conducive to comparative analysis according to different pore types,pore densities,mineral types,and dyeing conditions,and photograph them into images.Then use the method of median filtering to preprocess the selected image.Reduce the impact of noise and impurities in the image,eliminate interference factors that affect the analysis results,providing the basis for the analysis.Secondly,the components in the image are classified into four types: pores,strong erosions,weak erosions and rock skeletons.The interactive extraction method is used in turn to extract the training sample set and the test sample set,at the same time,set label type.The classifiers are trained by supervised support vector machines,Naive Bayes classifiers and unsupervised K-means clustering algorithms in machine learning respectively,then all the images are analyzed and identified.The experimental results show that SVM is sensitive to the selection of training samples,and it is not easy to control.The recognition of weak corrosion and rock skeleton is not obvious.K-means clustering algorithm due to unsupervised characteristics,it’s ability to recognize complex pores is not very well;Naive Bayes classifier in the most stable in the classification of such rock thin section image samples and with the highest recognition accuracy.Therefore,the analysis of pores adopt by the recognition results of Bayesian classifiers,the pores of each image were extracted independently,and the information such as the number of pores,the perimeter of the pores,the area of the pores were further obtained and pore parameters were calculated.