Application of YOLO Algorithm for Segmentation and Classification of Minerals in CT Slices Obtained by Dual- and Multi-Energy CT

This work addresses a problem of detecting minerals hidden in waste rock. Beryl, plagioclase, fluorite, and diorite particles were selected as target objects; host rock was muscovite. These substances have similar properties, therefore it is impossible to perform segmentation and classification using single projection X-ray images. For high-contrast distinguishing of beryl, fluorite, and plagioclase against the muscovite background, CT slice dual-energy transformation was used. To identify diorite particles and distinguish them from muscovite, multi-energy CT technique was used. In order to obtain different energy levels, metal filters were used in combination with voltage control at the pulsed X-ray source. Peak voltage of the source reached 450 kV. Segmentation and classification of the target particles in dual- and multi-energy slices were accomplished through the use of convolutional neural networks. In solving these problems, the YOLO algorithm demonstrated high accuracy and speed.