| In the field of oil and gas development,the component analysis of underground minerals and the identification of pores have always been the focus of research.Compared with geophysical methods such as indirect evaluation and characterization methods such as logging methods,rock slices based on core sampling and grinding can intuitively reflect the true distribution of underground minerals and have a more accurate description effect.Traditional thin slice identification is mainly performed by professionals through manual observation,which requires a lot of time and labor cost.It is easily affected by the researcher’s experience factors,which makes the level of identification results vary.Building an automated rock slice image segmentation and recognition system based on deep learning helps to save manpower and speed up the identification speed,which can obtain a more accurate description of underground distribution.Therefore,it is of great significance to study the segmentation and recognition technology of rock slice image based on deep learning.In this thesis,orthogonal polarized light sequence images of rock slices are used as experimental data,and the research idea of segmentation and recognition is determined.We mainly conduct research from deep learning segmentation models and machine learning mineral particle identification strategies.The main research contents are:(1)This thesis studies the structural characteristics,extinction characteristics and color characteristics of the image,which comprehensively depicts the rock slice sequence image from multiple angles.Taking the edge of the slice image as the target,using traditional edge segmentation and threshold segmentation methods(Canny algorithm,Sobel algorithm and adaptive threshold segmentation algorithm)to sequentially process the image sequence,and superimpose the segmentation results of multiple single images to obtain more Continuous edge distribution.Fully convolutional neural networks,deep convolutional autoencoders and U-Net networks are used as deep learning comparison models to compare and analyze the differences in segmentation effects of different structures.(2)This thesis studies the segmentation effect of different deep learning optimization strategies on rock slice images.With U-Net neural network as the backbone,Ms U-Net is proposed based on the Inception structure to enhance the multi-scale feature extraction ability of the model downsampling process.A Res U-Net based on the residual structure is proposed to alleviate the gradient missing problem caused by the depth of the model.Based on the dense connection idea,DU-Net is proposed.By adding feature sparse bypass connections,different levels of features can be integrated to enhance the generalization ability of the model.Combining the attention mechanism to construct DAU-Net,the network can recalibrate the channel weights for network-level fusion,which improves the overall learning strategy of the network.Various statistical indicators are used to construct an objective segmentation effect evaluation system.The experimental results show that under the same network hyperparameter settings,DAU-Net can obtain the best segmentation recognition effect.(3)According to the extinction characteristics of mineral particles,I propose a mineral particle identification method based on spatial statistics.Different types of machine learning algorithms are used to construct mineral particle lithology discriminant models to study and analyze the difference information of the inherent characteristics of different lithologies.Combining the results of deep learning segmentation,I propose a method for identifying lithology of mineral particles based on regional statistics.Experimental results show that this method can effectively smooth the recognition result image and better recognize the lithology and pores of mineral particles. |
