首页|Deep convolutional neural network for 3D mineral identification and liberation analysis
Deep convolutional neural network for 3D mineral identification and liberation analysis
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
Mineral liberation analysis (MLA) is an automated mineral analysis system that identifies minerals in polished two-dimensional (2D) sections of drill or lump cores or particulate mineral matter. MLA allows a wide range of mineral characteristics to be investigated, including fragment size, mineral abundance, and liberation. To date, this analysis has been primarily limited to two-dimensional (2D) section information. In this study, we describe an MLA workflow that enables the extension of MLA into 3D via the utilization of 3D X-ray microcomputed tomography and convolutional neural network (CNN) guided by M4-Tornado micro-X-ray fluorescence (microXRF) data. With the combination of 3D greyscale micro-CT data with several 2D identified element mappings, the state-of-the-art CNN architecture called EfficientU-Net-b3 is trained and tested for multimineral segmentation on both an intact complex iron ore sample and the corresponding crushed fragments. Compared to traditional manual segmentation methods, where only greyscale thresholds are selected by humans, CNN-based segmentation takes the information from unbiased microXRF and extracts not only the greyscale values but also the texture features from the image. After the segmentation of the 3D micro-CT datasets, several mineral liberation analyses are performed in the 3D domain, as well as 2D slices that are uniformly selected from the 3D segmented fragments data. The results from 2D and 3D MLA demonstrate that the 2D analysis results are heterogeneous and significantly different (up to a 14 % difference in the association indicator matrix) from the 3D analysis results. The loss of mineral information from 2D could influence ore body characterization and the proposed mineral processing procedure. Overall, the proposed workflow provides a digital mineral framework for 3D MLA for future ore characterization applications.
Deep LearningDigital imaging processing3D mineral identification3D liberation analysisX-ray microcomputed tomographyRAY MICRO-TOMOGRAPHYENERGY TRANSITIONSURFACE-AREAXRF
Tang, Kunning、Da Wang, Ying、Mostaghimi, Peyman、Knackstedt, Mark、Hargrave, Chad、Armstrong, Ryan T.
NSTL
Elsevier
