Multiple Kernel Learning Approximation Algorithm And Its Application In Tight Sandstone Pore Segmentation

Multiple kernel learning shows good generalization and flexibility when solving irregular and heterogeneous large-scale nonlinear data problems.The multiple kernel learning regularization path is a measure to solve multiple kernel learning many times and select the optimal model.The kernel matrix size is large,the computational cost is high and the efficiency of the optimization model is affected when multiple kernel learning regularization path processes large-scale data.How to simplify the calculation of matrix and improve the efficiency of regularization path is a problem worth studying.In the study of CT images of tight sandstone,the method of machine learning is used for pore and non-pore segmentation,which reduces the requirements on the professionalism of the researchers and can improve the efficiency of segmentation.The research work of this paper is carried out by matrix approximation to improve algorithm efficiency,image multi-feature fusion and pore segmentation.The main contents are as follows:1.A multiple kernel learning regularization path approximation algorithm(MKLRP_CUR)based on CUR matrix decomposition is proposed.The CUR matrix decomposition method is applied to the multiple kernel learning regularization path algorithm.Sampling technique is applied in solving multiple kernel learning regularization path.According to the probability of the rows and columns of the matrix,some rows or columns of the large-scale high-dimensional matrix are extracted.The high-dimensional matrix is decomposed into three low-rank matrices,and the product of three low-rank decomposition matrices is used to approximately replace the kernel matrix.Adjust the calculation order of all matrices to optimize the calculation efficiency of the algorithm.MKLRP_CURalgorithm reduces the scale of matrix computation,reduces the complexity of matrix computation,and improves the computational efficiency of multiple kernel learning regularization path algorithm.2.Feature extraction of pixel points of CT images of tight sandstone.In order to perform pore segmentation on CT images of tight sandstone,firstly,the image is preprocessed by bilateral filtering and histogram equalization.Then,the characteristics of pores and non-pores in the image are analyzed,and the color feature,texture feature,shape feature,edge feature and gradient feature of pixel points in the CT image are determined to prepare for the next step of pixel classification.3.Segmentation of pores in CT images of tight sandstone.Firstly,the local area containing pores in the sandstone CT image is intercepted,and the pixel sample set of the image is constructed.Then,a variety of features of pixels are extracted,and a classification model is established by using MKLRP_CURalgorithm.The input of the model is the feature extracted from the image pixels,and the output is the classification result of whether the image pixel is a pore.Finally,the trained classification model is used to classify the pixels of the whole CT image.According to the classification results,the pixel is binarized to generate the pore segmentation image.Bilateral filtering and morphological operations are used to post-process the generated pore segmentation images,so as to accurately segment the sandstone CT images.