Three-dimensional quantitative analysis on granular particle shape using convolutional neural network

To identify all desired shape parameters of granular particles with less computational cost, this study proposes a three-dimensional convolutional neural network (3D-CNN) based model. Datasets are made of 100 ballast and 100 Fujian sand particles, and the shape parameters (i.e., aspect ratio, roundness, sphericity, and convexity) obtained by conventional methods are used to label all particles. For the model training, by feeding the slice images of particles into the model, the contour of particles is automatically extracted, thereby the shape parameters can be learned by the model. Thereafter, the model is applied to predict shape parameters of new particles as model testing. All results indicate the model trained based on slice images cut from three orthogonal planes presents the highest prediction accuracy with an error of less than 10%. Meanwhile, the accuracy for concave and angular particles can be guaranteed. The rotation-equivariant of the model is confirmed, in which the predicted values of shape parameters are roughly independent of orientations of the particle when cutting slice images. Superior to conventional methods, all desirable shape parameters can be obtained by one unified 3D-CNN model and its prediction is independent of particle complexity and the number of triangular facets, thus saving computation cost.