Deep Learning and Pseudo-Labeling for Ore Granulometry

Abstract

Determining the fractional composition of ore using artificial intelligence is an urgent task in recent decades. Classical methods do not achieve decent quality and deep learning methods require huge labeled datasets. We propose an algorithm for automating the masking of ore stones boundaries on a conveyor using classical computer vision methods. Also, we propose a method for tracking the ore particle size distribution on the conveyor based on the U-Net neural network that was trained on the dataset utilizing pseudo-labels obtained with the algorithm. We compared several deep learning models for segmentation. Experiments on our dataset have shown that it is possible to segment ore with an accuracy of up to 70%. Also, we identified typical problems of granulometric analysis approaches on two-dimensional images and discussed ways for further possible development.