Super-Resolution Dataset of SEM Images: Generating Realistic Paired Images for Digital Cores

ABSTRACT: The digital core imaging technique is crucial for analyzing rock parameters. Super-resolution methods are indispensable for improving the resolution of digital core images. Because of factors like environmental noise, alignment challenges, and high scanning costs, these methods often rely on the bicubic interpolation technique to create coarse-resolution images, which may not accurately reflect the true coarse-resolution data. To overcome this limitation, two types of scanning electron microscopes (SEM) with magnifications of 1, 2, 4, 8, and 16 are employed to capture both high- and coarse-resolution images. We then train a residual network to model the residual mapping between these images. Using six algorithms - Lanczos3, Lanczos2, bilinear, bicubic, box and nearest - we generate coarse-resolution images from the fine-resolution scans. The residual network is subsequently used to reconstruct fine-resolution images from the six sets of coarse-resolution inputs. To quantitatively assess the performance of the reconstruction, we compare the Structural Similarity Index Measure and Peak Signal-to-Noise Ratio between the reconstructed images and the original fine-resolution images. This work presents a new method that overcomes the limitations of current image super-resolution techniques and addresses the issue of inaccurate coarse-resolution images generated by conventional methods.