Detection and Segmentation of Iron Ore Green Pellets Using Computational Vision

Resumo

In the industry, iron ore pellets, agglomerates with a diameter ranging from 6 to 16mm, composed
mainly of fine iron oxide particles, are one of the essential inputs used in the global production of steel, where
sphericity and strength of the pellets are necessary for the process, as well as the correct diameter, called the
particle size range. To provide mechanical resistance to the newly formed pellets, a characteristic that prevents the
pellets from breaking and turning into fines, a firing process is carried out where the thermal efficiency of this
pellet is intrinsically linked to the ideal diameter and humidity of the pellets. In the work, two neural models of
deep learning were presented and compared among themselves in the segmenting, and then measuring the diameter
of each of the iron ore pellets, they are Mask R-CNN and YOLACT. Such work makes possible improvements in
the controllers of the pelletizing discs, improving the quality of the pellets as a whole, as well as, a greater precision
in the desired granulometric range of the pellets. It was seen that the two networks mentioned had excellent results,
however, the Mask R-CNN proved to be more costly in processing compared to YOLACT.