Automated Extraction of Sediment Core and Scale Segments from Core Scanner Images

In this paper we report on semantic segmentation of images of sediment cores obtained from core scanners. Such images usually encompass several objects beyond the essential parts showing the core itself and the measuring scale (typically a ruler or some sort of tape measure), which might be necessary for imaging or physically holding the material, but are irrelevant for the analysis of cores. Moreover, such objects are obstacles to analyzing cores in an automated manner by image processing or deep learning methods. As part of development of a toolkit for automated analysis of sediment cores, we show that the essential parts of the images - core itself and scale reference - can be extracted with high accuracy, using well-established convolutional architectures for semantic segmentation, such as Unet.