10:45 AM - 12:15 PM
[SCG58-P08] Data-Driven Super-Resolution for Rock Sample CT Images Based on Sparse Modeling: Applications to the Complex Texture of Serpentinite
Keywords:Oman Drilling Project, image superresolution, sparse representation
In recent years, imaging of rock structures using an X-ray CT scanner has been widely used for analysis in scientific drilling. Since rocks have structures of various scales, it is necessary to acquire and analyze the structure of each scale from the imaging data. However, the resolution of the X-ray CT used in drillships is rather low, limiting the ability to acquire the fine structure of the rock, so the CT data is not fully utilized. On the other hand, in the case of imaging using high-resolution X-ray CT used in places such as laboratories, although it is possible to acquire fine structures, the range and number of images that can be acquired is limited. Therefore, it is important to increase the resolution of X-ray CT data acquired on drillships by utilizing the limited data captured by high-resolution X-ray CT scanners.
In this study, we propose a super-resolution technique using sparse representation and dictionary learning for CT images of rocks. Using the super-resolution using sparse representation, we estimate dictionary for basis images which reflect characteristics of spatial structures in rock samples. The sample applied in this study is a serpentinized dunite produced at various stages of hydrothermal alteration in the crust-mantle transition zone of the Oman ophiolite. The sample consists mainly of serpentinite, hydrolite, magnetite, and Cr-rich spinel. In the proposed method, a high-resolution dictionary is first learned by extracting the rock structure using sparse representation. A low-resolution dictionary is also created by down-sampling the high-resolution dictionary. By assuming common sparse representation between high- and low resolutions, we estimate the reconstructed image. In particular, we construct data-driven framework for three-dimensional super-resolution of rock CT images in addition to the framework for two-dimensional super-resolution.
By applying the proposed method to the rock CT images, we show that the estimated high-resolution images applying the proposed method reconstructed microstructures such as spinel, magnetite, and mesh structure. Comparison with conventional interpolation methods using multiple evaluation indices also showed the superiority of the proposed method.
In this study, we propose a super-resolution technique using sparse representation and dictionary learning for CT images of rocks. Using the super-resolution using sparse representation, we estimate dictionary for basis images which reflect characteristics of spatial structures in rock samples. The sample applied in this study is a serpentinized dunite produced at various stages of hydrothermal alteration in the crust-mantle transition zone of the Oman ophiolite. The sample consists mainly of serpentinite, hydrolite, magnetite, and Cr-rich spinel. In the proposed method, a high-resolution dictionary is first learned by extracting the rock structure using sparse representation. A low-resolution dictionary is also created by down-sampling the high-resolution dictionary. By assuming common sparse representation between high- and low resolutions, we estimate the reconstructed image. In particular, we construct data-driven framework for three-dimensional super-resolution of rock CT images in addition to the framework for two-dimensional super-resolution.
By applying the proposed method to the rock CT images, we show that the estimated high-resolution images applying the proposed method reconstructed microstructures such as spinel, magnetite, and mesh structure. Comparison with conventional interpolation methods using multiple evaluation indices also showed the superiority of the proposed method.