Investigation of three-dimensional aggregate contact evolution using an enhanced image segmentation algorithm

In aggregate skeleton structure research, the distribution of aggregate contact points is a key factor affecting structural stability under loading. However, traditional image segmentation methods struggle to accurately segment aggregates in CT images, making the study of aggregate contact in three-dimensional space in real specimens a challenge. To address this, this study first proposes an adaptive bilateral filtering algorithm for CT image preprocessing to enhance edge details. Then, an improved U-Net algorithm is proposed, which combines Inception convolution modules, residual connections, spatial attention mechanisms, and a novel loss function incorporating boundary information, effectively solving the adhesion issue in aggregate segmentation. Finally, based on the centroids and spatial orientation of aggregates, an approximate ellipsoid model is constructed, and the contact number of the main skeleton aggregates is calculated, further analyzing the evolution of contact numbers during loading.