Analyzing the pore structure of pervious concrete based on the deep learning framework of Mask R-CNN

Abstract

Analyzing the pore structure and establishing the relationship between pore structure and macroscopic property are important research topics for pervious concrete. However, the current method of pore structure analysis is manual threshold segmentation, and its analysis efficiency is low and accuracy needs to be improved. The purpose of this paper is to establish an efficient and accurate method for analyzing the pore structure of pervious concrete based on deep learning. The pervious concrete CT slices were used as the dataset, and the improved Mask R-CNN algorithm was used as the model training framework to train the pore recognition model. The influence of the improved algorithm on the pore identification effect was analyzed. The effectiveness of the pore structure identification and analysis of the training model in this paper was evaluated compared to the traditional methods i.e., manual threshold segmentation and watershed algorithm. The results show that the improved algorithm shows a better convergence and more accurate in pore segmentation and identification compared to Mask R-CNN. As control groups, manual threshold segmentation is prone to under-segmentation or over-segmentation, watershed algorithm will cause over-segmentation or mis-segmentation. In contrast, the pore identification results of the model in this paper are closest to the ground truth, and there are few over-segmentation, under-segmentation or mis-segmentation. Additionally, the model also has good robustness to the brightness and contrast of the CT slices.