Key Technologies Of Three-dimensional Reconstruction Of Core Imagery Using Gray Scale Images

The microstructure of the three-dimensional(3D)core has an important influence on its physical properties.With the help of the 3D images,the microstructure of core and its physical properties can be analyzed more effectively.Mathematical modeling is one of the main methods to obtain 3D core microstructures.Most of the existing researches on 3D core structure based on mathematical modeling focus on binary image,while core 3D images in gray level can provide more information for characteristic analysis.Thus,the methods of 3D reconstruction of core imagery using gray scale images was studied in this disseration.Firstly,to characterize the similarity of the 3D core image in gray level,a similarity measurement method for gray level core images was proposed.Secondly,a 3D super-resolution algorithm based on self-learning for 3D rock images and a 3D super-resolution algorithm based on Very Deep Convolutional Networks superresolution(VSDR)were proposed.Thirdly,a 3D reconstruction method from a single 2D gray image was proposed.Finally,a multi-scale fusion method for core images was proposed.The main research contents and innovations of this paper include the following four aspects:(1)The shortcomings of the existing model similarity calculation methods are analysed and proved by experiments.A more suitable method of pattern similarity calculation method was proposed.Besides that,a core image similarity evaluation algorithm was proposed based on the method of pattern similarity calculation method.Further more,an adaptive class setting method is proposed.Finally,the algorithm is combined with the similarity evaluation method of the two-phase core image to make up for the deficiency of the existing similarity evaluation method in evaluating the similarity of the gray-scale core image,and realizes the joint evaluation of the similarity of the gray-scale core image from a number of characteristics..(2)In order to improve the quality of 3D core grayscale images and lay a foundation for subsequent multi-scale core CT image fusion reconstruction,the super-resolution reconstruction algorithm of 3D images of core based on selftraining and sparse representation and the super-resolution reconstruction algorithm of 3D image of core based on deep learning are studied in depth.According to the characteristics of 3D core image,in order to improve the quality of core image,an appropriate noise suppression algorithm and three-dimensional feature extraction method are introduced when building an overcomplete dictionary.Theoretical analysis and experimental verification of the reconstruction parameters are carried out.In the results of comparative experiments,the proposed algorithm is improved by an average of 1.553 on PSNR and 0.02 on SSIM.In addition,the cross-correlation function was used to optimize the set of eigenvectors to reduce the reconstruction time.In order to further improve the effect of super-resolution reconstruction,a 3D core CT image super-resolution reconstruction algorithm based on deep learning was proposed.Besides that,the residual network model is improved to alleviate the problem of gradient disappearance.Further more,a new cost function calculation method combining mean square deviation and SSIM was proposed.The experimental results show that compared with the self training and sparse representation-based core 3D image super-resolution reconstruction algorithm,the improved algorithm based on VDSR has an average improvement of 1.57 in PSNR and 0.022 in SSIM;Compared with the original VDSR algorithm,the average improvement is 0.51 in PSNR and 0.007 in SSIM.(3)In order to solve the problem that most of these algorithms can only reconstruct the binary image,the 3D reconstruction algorithm based on 2D core image is deeply studied,,a 3D reconstruction algorithm based on a single 2D gray core image was proposed.According to the prior information of strong correlation between two adjacent layers of the core sequence images,a fast-sampling algorithm was proposed.With this method,the continuity and variability between adjacent layers could be constrainted during the reconstruction process.A reconstruction strategy of central area simulation was proposed to reduce the blocking effect.The 3D gray core image reconstructed by this algorithm could not only reflect the distribution of pores in three-dimensional space,but also provide more gray value and texture information,which was helpful to further study the distribution of core density and composition.The experimental results showed that the reconstructed image and the reference image are not only similar in visual effect,but also very similar in two-point probability,linear path,local porosity,gray histogram and pattern classification function.It also indicates that the proposed reconstruction algorithm could reproduce the multiple features and pattern structure information of 2D gray image in 3D structure,and effectively reconstruct the 3D core structure with complex morphological features similar to the training image.(4)A multi-scale fusion algorithm for the core CT images was proposed based on super-resolution reconstruction method and 3D core reconstruction method to alleviate the problem that the image of large core sample acquired by imaging equipment is not clear enough and cannot effectively describe the pore structure of small size,while the image of small core sample was not globally representative enough.With our multi-scale fusion algorithm,the low-resolution core image was used to provide global feature information for the fusion image,such as the basic morphological structure of large-scale pore and particle,and then the texture pattern and the small pore distribution information of high-resolution core image were used to refine the fusion image.In order to estimate the similarity of core images in different scales,a blind image quality assessment was involved.In addition,a multilevel pattern mapping dictionary containing local binary patterns was designed to speed up the pattern matching procedure.Meanwhile,an adaptive weighted reconstruction algorithm was designed to work together with the pattern similarity calculation method proposed to reduce the block effect in this dissertation.Experimental results showed that the proposed algorithm can realize the fusion of CT images in multiple scales of the same core sample well.The fused image maintained not only the long-range information and the overall structure information of the large-scale image,but also the small pore information and texture information in the high-resolution image.