A quantified study of segmentation techniques on synthetic geological XRM and FIB-SEM images

Three sets of synthetic images were created from two original datasets. A suite exhibiting greyscale contrast was produced from an 8.96-mu m voxel size 3D X-ray microscopy image of a sandstone rock and a two suites (one showing greyscale contrast and one showing both greyscale and textural contrast) were produced from a 5 x 5 x 5 nm voxel size FIB-SEM image of a shale rock. The performance of three image segmentation algorithms (global multi-Otsu thresholding, seeded watershed region growing, and machine learning-based multivariant classification) was then assessed by their ability to recover their respective original segmented 3D images. While all algorithms performed well at low noise levels, machine learning-based classification proved significantly more noise tolerant than either of the traditional algorithms. It was also able to segment the non-greyscale (textural based) contrast, something the traditional completely failed to do, with voxel misclassification rates for the traditional techniques above 50% at a 0 noise level within the textural contrast regions. Machine learning-based classification, in contrast, achieved misclassification rates of less than 5% in the same regions.