Computing Particle Size Distribution of Mineral Rocks using Deep Learning-based Instance Segmentation

This work presents a deep learning based system for estimating the particle size distribution of two types of rocks in a flow-through environment in mineral processing. Deep Learning has become one of the most important topics in Computer Vision, however, less is known about applying deep neural networks in industrial use cases, where data availability is very limited. Due to this limitation, previous works have focused the efforts on generating synthetic images and benchmark against them or against similar datasets. Because slurry environments exhibit a high level of complexity and image noises, it is almost impossible to transfer knowledge from synthetic data, hence our efforts are aimed towards working only with real data. Target images contain apatite and phlogopite particles in slurry, presenting complex scenarios like overlapping particles, blurry or mixed particles. The proposed system segments all instances of particles, then computes a size distribution based on the predicted masks and the magnification of the imaging device. Finally, the result is benchmarked against a reference method, that uses laser diffraction. Two state-of-the-art neural networks are compared, highlighting tradeoffs that need to be considered for a practical implementation of the system.