Application of Deep Convolutional Neural Networks for Analysis of Appar- ent Density and Porosity of Iron Ore Pellets

Resumo

The porosity and apparent density of iron ore pellets directly interfere with the blast furnace process
and, therefore, need to be known to assist in its control and optimization. These characteristics are generally
calculated using a pycnometer that uses mercury under pressure to fill the pores of the pellet. Considering the
need to preserve the environment and the safety of operators, proposals were made to replace this process, but
there are several complaints about the repeatability of results achieved, in addition to the time spent in preparing
and executing these essays. At the same time, it is possible to observe a remarkable development in Computer
Vision and Artificial Intelligence, mainly through Convolutional Neural Networks, which can extract patterns
from a set of images and detect these same patterns in images subsequently exposed to this network. In addition
to performing classification and detection, the Mask R-CNN network can perform pixel-by-pixel segmentation of
objects in images. In its evaluation, the network presented a significantly high mAP and accuracy, demonstrating a
satisfactory result for the segmentation and obtaining of porosity and apparent density values, with results similar
to the essays currently used.