Automatic identification method for three-phase structure of pervious concrete based on deep learning network of Mask R-CNN

Pervious concrete is a three-phase structure, including the pore phase, aggregate phase and paste phase. Currently, the pore phase of pervious concrete can easily be identified with CT and image processing technology. However, there are still technical limitations to identifying the aggregate and paste phases. This paper aims to establish a three-phase structure analysis method for pervious concrete based on deep learning. 30 images of the cross section of nine groups of color-pervious concrete specimens with different coarse aggregate sizes were obtained to form the initial dataset. The initial dataset was preprocessed, labeled and extended, and four deep learning frameworks were selected to train the three-phase structure identification models. By comparing and analyzing the identification effects of different models, the model suitable for three-phase structure identification was determined. The segmentation accuracy of each phase was evaluated. The results show that the method for preparing the dataset required for three-phase structure identification model training could improve identification accuracy. The model can accurately identify the three-phase structure with fewer false and missed identifications. The descending order of three-phase structure identification accuracy was pore phase＞aggregate phase＞paste phase. This was related to the quality of the dataset and annotation accuracy.