Style Transfer applied to CT image downscaling: a study case from Brazilian Coquinas

摘要

The identification of micropore systems in carbonate rocks is an important task of image processing because of the high impact these systems cause on fluid flow. Currently, one of the main tools used to characterize rock samples is computed tomography (CT). Such micro information poses the challenge associated with the limitation of the CT's resolution. Therefore, we propose an alternative method of inserting the micropore features of mu CT to lower resolution images, but with higher coverage. We can perform this by a novel application of Style Transfer that can insert the heterogeneity pattern of high-resolution (HR) images (CT of 7 and 40 mu m resolution) into low-resolution (LR) images (CT of 90 mu m resolution), downscaling the image through a super-resolution method. This technique uses the power of VGG19, a convolutional neural network that won the ImageNet Large-Scale Visual Recognition Challenge in 2014, as a texture extractor. We applied this novel technique to condensed shell rocks, called coquinas, from the Itapema Formation, Santos Basin offshore Brazil. The porosity of the LR image, with initial average value of 11%, resulted in an average porosity of 12% (40 mu m res.) and 13% (7 mu m res.) after downscaling. This is closer to the porosity range of the coquina (13% to 32%, with a mean of 21%) and an increase in the porosity of 6% and 19% in average, respectively. Despite this, the connectivity of the original LR CT was of 3% on average and, in the simulated HR CT, the connectivity was of 5% (40 mu m res.) and 6% (7 mu m res.). In addition, in such examples, this method inserted connectivity in directions that were null in low-resolution images before the style transfer. Hence, the results demonstrated that the Style Transfer offers an alternative for downscaling CT images by inserting the texture from high-resolution images.