Segmenting Low-Contrast XCTs of Concretes: An Unsupervised Approach

A research team has published a paper titled 'Segmenting Low-Contrast XCTs of Concretes: An Unsupervised Approach' (arXiv:2603.00127), presenting a novel method to train a convolutional neural network (CNN) for semantic segmentation without manually labeled data. The core challenge addressed is the low contrast in X-ray computed tomography (XCT) scans of concrete, where aggregates and mortar have similar X-ray attenuation coefficients, making traditional analysis difficult and labeling expensive. The proposed solution is a self-annotation technique that leverages superpixel algorithms to identify perceptually similar regions and relates them to the global image context using the CNN's receptive field.

The technical approach enables the model to autonomously learn the 'global-local relationship' within the images, identifying semantically similar structures like aggregates and cement paste. This unsupervised methodology represents a significant shift for materials science and non-destructive testing, where acquiring large, expertly labeled datasets for every new concrete mix is a major bottleneck. The work opens avenues for more accessible and scalable analysis of concrete's internal microstructure, which is critical for assessing durability, strength, and failure mechanisms. The authors discuss the model's performance on their XCT datasets and outline potential paths for further improvement in this promising research direction.