Digital Rock Core Images Resolution Enhancement with Improved Super Resolution Convolutional Neural Networks

Abstract

It has proved that Super Resolution Convolutional Neural Networks (SRCNN) can enhance the resolution of images by other researchers. We applied SRCNN to enhance Digital Rock Core Images that play an important role in analyzing rock core in petroleum industry. We noticed that there are a large number of similarly colored patches in Digital Rock Core Images. Then we realized if we applied the original SRCNN to our project directly, we would waste some time in the first step of the SRCNN which is to establish a data set with the bicubic interpolation algorithm. This is because the results are almost the same if we use the bilinear interpolation algorithm to replace the bicubic interpolation algorithm to deal with pixels in the similarly colored patches, and the time complexity of bilinear interpolation algorithm is smaller than that of the bicubic interpolation algorithm. In order to save time, we proposed an improved SRCNN and tested the performance with 2760 Digital Rock Core Images. Experiments show that the improved SRCNN reduces 11.1% of computation time for preparing data set, but the performance of Super Resolution is similar to the original SRCNN. The proposed method can save many electricity bills when we deal with millions Digital Rock Core Images for future job. In addition, it has practical value for Super Resolution tasks with SRCNN in other fields.