On the tracking of shelly carbonate sands using deep learning

It is well known that carbonate sands possess weak mineralogies, complex particle morphologies and porous microstructures. These characteristics lead to very distinct mechanical properties of carbonate sands, such as low shear strength, high crushability and high permeability. This paper presents a novel investigation of the recognition and tracking of intact carbonate sand particles using a deep learning method called PointNet++. The capability of PointNet++ to extract the global and local features of the porous structures of carbonate sand particles enables it to excel in the pattern recognition of porous granular materials. First, for the reconstruction of carbonate sand particles, a set of two-dimensional raw images obtained from X-ray microtomography scanning were handled by a series of image processing techniques such as median filter, segmentation and thresholding algorithms. In particular, a special technique previously developed by the authors was used to treat the abundant intra-particle pores and surface concavities of carbonate sand particles to avoid the problem of image over-segmentation. Second, to prepare the training datasets to be used in the PointNet++ deep learning exercise, a strategy of sampling and grouping was proposed to divide the initial point set of each sand particle into several groups. Next, PointNet++ was utilised to capture the global and local context features of the sand particles at different length scales and shown to successfully recognise and track all of the particles. Finally, a comprehensive comparison between several particle tracking methods reported in the literature was made, and the outstanding advantages of the deep learning-based particle tracking method were summarised.