Evaluation of the particle size distribution of on-site rockfill using mask R-CNN deep learning model

Particle size distribution (PSD) of the on-site rockfill is one of the most critical factors in evaluating the quality assessment of the rock-filled concrete (RFC). Due to the large quantities and volume, it is difficult to measure the grain size of each rock manually. Image-based methods are widely adopted for the grain segmentation, but the result is not ideal if the rocks are closely connected and overlapped. In this study, a new model Mask R-CNN from the perspective of deep learning is deployed to develop an automatic measurement method of rockfill PSD. The model training was conducted using photos captured from on-site rockfill in Fengguang RFC dam of China. The results of the trained model agree well with the artificial measurements, and it proves Mask R-CNN as an effective technology for the automated estimation of the rockfill PSD in the engineering practice.