Multi-scale Fusion Super-resolution Reconstruction Of Core CT Images Based On Deep Learning

With the continuous development of science and technology,the demand and importance of high-resolution images are gradually increasing.Compared with low resolution images,high resolution images contain more pixel information and can meet the needs of military,medical,public security and other fields.The application of deep learning method in the field of superresolution reconstruction of rock imaging can provide more rock images and more detailed rock information,which not only helps researchers to understand the microstructure of rock mass,but also lays a solid foundation for the development and utilization of oil and gas resources.CT technology has the advantages of nondestructive,three-dimensional and computer integration,and has been widely used in reservoir geology research.However,at present,limited by the hardware conditions of detection equipment,imaging field of view and imaging resolution are often compromised,so that the resolution of the image is low.In order to obtain higher resolution and more accurate data,it is necessary to use super resolution image reconstruction technology.The traditional low resolution imaging method has some limitations in feature extraction,which cannot recover the high frequency information of oil and gas reservoir well,and lacks the key fine texture information.In view of this,this paper takes multiscale fusion,residual structure,U-Net model and generative adversarial network as the basis.On this basis,we carry out the research on the super-resolution reconstruction of core CT images,study the multi-source image data fusion algorithm based on deep learning,and improve the reconstruction speed on the premise of guaranteeing the reconstruction quality.The main work of this thesis includes the following aspects:(1)A multi-scale fusion super-resolution reconstruction algorithm of core CT images based on residual U-Net is proposed(MS-Res Unet).In this study,a residual channel is adopted to overcome the problem of image blurring caused by insufficient processing of high frequency information and information missing.At the same time,the algorithm is combined with multiscale fusion technology to obtain high resolution image.In order to better preserve the information of the image,the method of up and down sampling is adopted.Moreover,multiresolution image sub-network is used in parallel to realize multi-scale information fusion.The proposed deep neural network based on multi-scale fusion can better excavate the nonlinear mapping relationship between high,medium and low-resolution images.(2)A multi-scale fusion super-resolution reconstruction algorithm for core CT images(MS-Unet GAN)by integrating multi-scale U-Net and GAN is proposed.In this study,two UNet structure with excellent performance are used as generator and discriminator,respectively.Meanwhile,the feature information of three scales is input simultaneously at the input end and fused.Thus,the model’s ability to restore image details is improved.Then,the feature processing and optimization of the image are carried out through U-Net’s up-and downsampling and skip connection,and high-level feature information is extracted from the lowresolution image.GAN is used for extracting textural feature information to generate better high-resolution images.The proposed model is capable of generating high-resolution images with better detailed textures with higher accuracy and fidelity through multi-scale feature fusion and GAN optimization.