Research On Pore Extraction And Granule Segmentation Of Rock Cast Slice Image

In the petroleum geology industry,the traditional method to study thin sections of rock casting is to use a microscope to observe such thin sections,mark the casting area or particles,and then carry out statistical analysis.This method requires high operating experience of researchers.The work of casting or particle marking is heavy,and it is easy to cause some sample marking errors due to the subjective reasons of the marking personnel.With the development of image processing technology,it is common for the petroleum geology industry to analyze the cast thin section with the help of digital image technology.The artificial interaction method is used to extract the pores of the image and segment the particles,which improves the analysis efficiency and accuracy of the rock cast thin section image.However,due to the color difference of different casting areas in the rock casting thin section,and the complexity of particle morphology and location distribution,the accuracy of manual interactive extraction cannot meet the specific requirements.The extraction of pore areas and the segmentation of rock particles are still the difficulties in the preliminary work of the rock casting thin section analysis.Based on in-depth learning,this paper further optimizes the segmentation results,from realizing the edge detection of rock particles and the automatic extraction of pore areas,which carries out particle segmentation based on the extracted slice image,especially in-depth slice analysis.The research contents mainly include four parts.1.This paper analyzes the basic theory of convolutional neural networks(CNN)briefly,analyzes the differences between CNN and traditional artificial neural network(ANN).The key links in the structure of CNN are discussed in detail,such as convolution operation,down sampling,etc.,which expounds the advantages of weight sharing,local connection in CNN.The data set used in network training is introduced,the image is segmented,and the image enhancement method is introduced to further expand the data set for improving the generalization ability of the model that may prevent the over fitting phenomenon from affecting the experimental results.2.Casting region extraction is based on full convolution neural network U-Net.The casting segmentation effect based on RGB and HLS color space are limited by the color distribution characteristics of data set samples.In this paper,the semantic segmentation of casting region is realized by using full convolution neural network.The training network model classifies the casting and background regions of the original casting slice image,and evaluates the prediction results of a variety of full convolution neural networks using four evaluation criteria.Finally,the network parameters and the training time required by the network in the rock casting image data set are compared,which show the advantages of U-Net network in casting segmentation.3.Edge extraction of rock particles is realized on deep learning RCF network.The traditional edge detection algorithm can also detect the internal texture of rock particles in the process of edge detection.In this case,detection results can not be applied to the task of rock particle segmentation.In this study,neural network is used to detect the edge of rock particles.After comparing different network models,the RCF network is finally selected.Through the training model,the image pyramid is constructed after multi-scale scaling and transformation of the image,and the image binarization is used to remove the false edge,so as to further improve the accuracy of rock particle edge detection.At last the detection results are evaluated,which proves the advantage of deep learning in edge detection.4.On the basis of using neural network to detect the edge of rock particles and extract the casting region,the rock particles are segmented.There is an inseparable relationship between rock edge detection and casting region segmentation,and the results of above two steps can be superimposed to achieve the effect of rock particle segmentation.In this paper,through morphological expansion and morphological corrosion,the phenomenon of erroneous segmentation at the edge resulting from superimposed particle edge prediction map and casting area prediction map is solved,and the particle segmentation of rock image is completed,which is compared with the traditional threshold segmentation results.The experimental results show that the rock particle segmentation based on depth learning is better than the traditional threshold segmentation.