Abstract: No obvious grayscale feature can be detected in the mesoscopic damage area of concrete CT images, and it is difficult to extract mesoscopic damage information with an image segmentation method that is based on thresholding or edge detection. This paper describes a new K-Means clustering algorithm for deeply excavating such information from the CT images of concrete. First, we conduct uniaxial static compression tests on concrete cylinder specimens. Then, we determine the number of optimal clusters according to the outline coefficient, use this algorithm to search for the optimal partition of a CT image in non-supervised state, and obtain a partition map that carries mesoscopic damage information. Finally, the degree of damage is calculated by counting the total number of pixels over the damage area. The results show that at each compression stage, the evolution of mesoscopic damage in concrete can be observed intuitively on the map of failure zones and mesoscopic damage zones. It reveals the trend of mesoscopic damage degree varying with stress ? a relatively stable initial period and a stable growing period before the load peak, and after that an unstable decaying period. Thus, this study demonstrates the significant advantage of the K-Means clustering algorithm in analysis of the damage evolution of concrete.

Key words: concrete, CT image, mesoscopic damage, K-Means clustering algorithm, partition map, degree of damage