Research And Application Of Rock Image Classification Based On Convolution Neural Network

Rock is a very complex porous media through a lot of research.Some factors have an direct impact on the permeability of the rock,porosity and others,such as rock type and particle size,pore structure,fracture ratio,mud content and so on.It can provide a scientific basis for the research,exploration and development of oil and gas reservoirs by analyzing the microstructure characteristics of the rock,and formatting an effective rock microstructure feature extraction and classification methods.It is of great practical significance to the development and research of oil and gas fields.This thesis focuses on deep machine learning method of the current hot spot,especially the convolution neural network,and studies the classification of rock images with the theory of discrete cosine transform and discrete wavelet transform,the main works and achievements are as follows:Firstly,a convolution neural network rock image classification framework is proposed by training,verifying and testing the convolution neural network under different training iterations.Based on this framework,it classifies the rock images,which are adjusted by brightness,some geometric transformations and noise,and obtain a certain classification accuracy.Secondly,the convolution neural network rock image classification framework based on the discrete cosine transform is proposed.Based on this framework,the original rock image and the rock image after brightness adjustment,some geometric transformation and adding noise are classified and tested.The classification error rate are lower than the previous classification framework,and the training time is about 0.46 times of the previous classification framework.Thirdly,the convolution neural network rock image classification framework based on the discrete wavelet transform and discrete cosine transform is proposed.Based on this framework,the original rock image and the rock images which are adjusted by brightness,some geometric transformations and noise added are classified and tested.The training time of this framework is about 0.31 times of the second framework and its classification error rates are about equal to the second framework.The convolution neural network rock image classification framework is the rock image classification framework with low training time and low classification error rate through thecomparison.It based on discrete wavelet transform and discrete cosine transform uses a small number of coefficients to represent the rock image,thus shortening the training time of the frame.And it transforms the rock image into the frequency domain to extract the effective feature,and makes the effective classification of the rock image based on this feature.