Automated Mineral Classification

A photon counting detector gives X-ray transmission radiographs of a slice of a sample in which transmission is resolved into 128 bins of X-ray energies from 20 keV to 160 keV. After plotting the graph of transmission over energy bins, the K-edge can be traced. By using machine learning and computer vision techniques on these ‘energy bins vs derivative of X-ray transmission’ information, slices were not only classified much faster in an automated way but also performed better when compared to the manual classification of minerals by using intensities or gray scale values of particles.
Machine learning was implemented on the slices of manually prepared sample containing gold and lead particles, Printed Circuit Board (PCB) and a rock sample. Slices were also classified by implementing machine learning on intensity properties of gold and galena to further confirm an advantage of using spectrum information. Results helped to understand the challenges in the project and thus paved a way for advanced research.