11:00 〜 13:00
[MGI34-P05] Superresolution of X-ray CT images of serpentinites by sparse modeling
キーワード：Oman Drilling Project、image superresolution、sparse representation
In recent years, the X-ray computed tomography (CT) scanner has been commonly used for the onboard cores analyses for scientific drilling, which provides us the 100 to 1000 m scale continuous data of various structures (lithological or sedimental layering, fractures, pores etc) of rocks. However, the resolution of X-ray CT of such cores (e.g. about 0.17 mm for X-ray CT of onboard CHIKYU) is rather low for revealing the shape and geometry of mineral grains and thin fractures. In contrast, the resolution of X-ray CT in a typical laboratory is high (~0.01 mm or less). Therefore, if it is possible to link high-resolution X-ray CT in the laboratory of a limited number of samples with low-resolution CT data of boring cores, valuable big data for the analysis of multi-scale structures can be obtained.
In this study, we applied sparse super-resolution techniques to rock CT samples. Sparse super-resolution, which employs a sparse coding framework in machine learning, provides an efficient representation for target images. The sample used in this study is a serpentinite with distinct antigorite veins in the Moho transition zone, taken by the Oman Drilling Project (CM1A). The X-ray CT images were taken with voxel sizes of 10, 20, 40 and 80 micrometres. In the sparse super-resolution technique, we assume that both low-resolution and high-resolution images are expressed as linear sums of a small number of respective basis images. The sparse coefficients are considered to be common regardless of image resolution. In the dictionary learning process, a set of basis images is obtained from a plurality of rock CT images. In the super-resolution process, the sparse representation is estimated from the low-resolution X-ray CT images by using the low-resolution dictionary, and the high-resolution image is estimated by using the common sparse coefficients and the high-resolution dictionary. In this study, sparse super-resolution was applied to X-ray CT images of rock core samples, and 4 × 4 super-resolution was performed. The X-ray CT high-resolution images estimated by sparse super-resolution showed fine structures such as spinel alteration, antigorite veins, and mesh structure. By comparing the proposed method with the conventional interpolation algorithms, the effectiveness of the proposed method was demonstrated by evaluating the image quality indices and distribution of pixel intensities.
In this study, we applied sparse super-resolution techniques to rock CT samples. Sparse super-resolution, which employs a sparse coding framework in machine learning, provides an efficient representation for target images. The sample used in this study is a serpentinite with distinct antigorite veins in the Moho transition zone, taken by the Oman Drilling Project (CM1A). The X-ray CT images were taken with voxel sizes of 10, 20, 40 and 80 micrometres. In the sparse super-resolution technique, we assume that both low-resolution and high-resolution images are expressed as linear sums of a small number of respective basis images. The sparse coefficients are considered to be common regardless of image resolution. In the dictionary learning process, a set of basis images is obtained from a plurality of rock CT images. In the super-resolution process, the sparse representation is estimated from the low-resolution X-ray CT images by using the low-resolution dictionary, and the high-resolution image is estimated by using the common sparse coefficients and the high-resolution dictionary. In this study, sparse super-resolution was applied to X-ray CT images of rock core samples, and 4 × 4 super-resolution was performed. The X-ray CT high-resolution images estimated by sparse super-resolution showed fine structures such as spinel alteration, antigorite veins, and mesh structure. By comparing the proposed method with the conventional interpolation algorithms, the effectiveness of the proposed method was demonstrated by evaluating the image quality indices and distribution of pixel intensities.