ECS-Net: An adaptive feature learning semantic segmentation network for accurate concrete porosity prediction

Porosity prediction is important for concrete engineering safety and quality, but reliable computed tomography (CT) image segmentation remains challenging because of evolving internal pore structures, tiny pores, low grayscale contrast, and intricate morphologies. This study develops efficient concrete segmentation network (ECS-Net), an encoder-decoder semantic segmentation network equipped with three dedicated core modules for adaptive pore feature extraction from complex CT backgrounds. Its multiscale encoder integrates global structural information with local tiny-pore details, while the complexity-aware decoder adaptively improves segmentation at fuzzy boundaries and in low-contrast regions. Experiments on a self-built concrete CT dataset, Crack500, and aggregate datasets show that ECS-Net achieves improved segmentation performance and generalization ability compared with mainstream models. Three-dimensional reconstruction experiments using a limited number of specimens further confirm its stable performance in porosity prediction. These findings suggest that ECS-Net can support nondestructive detection of internal damage in concrete, structural integrity diagnosis, and long-term performance evaluation of in-service concrete structures.