Abstract: Accurate segmentation of aggregates in a drilling image of hydraulic concrete is very important for evaluating hardened concrete segregation, but traditional aggregate segmentation methods suffer drawbacks in low accuracy and weak generalization capability. This paper develops an improved Mask R-CNN model for segmenting hardened concrete aggregates by introducing an efficient channel attention module (ECAM) and a stage attention module (SAM) into the backbone Mask R-CNN network, so that its convolutional network can adaptively adjust the weights of the channel and space, significantly improving its capability of detecting the boundary and location distribution of target aggregate. And we work out a new method for the segregation degree of hardened concrete, realizing quantitative evaluation by quantifying the area and boundary of target aggregate and its distribution over different elevation. A case study shows that our aggregate segmentation model has an average accuracy of 0.875, or 4.2% higher than that of Mask R-CNN. In various complicated applications, its segmentation effect is better than that of traditional aggregate segmentation, and its errors of segregation degree is only 4.9% on average, verifying its effectiveness and superiority as a new tool for hardening concrete evaluation.

Key words: hardened concrete, Mask R-CNN, aggregate segmentation, concrete segregation, quantitative evaluation