Discrimination of pores and cracks in iron ore pellets using deep learning neural networks

Abstract The discrimination between pores and cracks is an important step in the microstructural analysis of iron ore pellets. While the porosity is fundamental during the reduction process in blast furnaces, cracks are strongly detrimental to the mechanical strength. The usual image processing tools cannot automatically discriminate between these two types of features, especially in 3D images obtained, for instance, with x-ray microtomography (microCT). As pores and cracks have essentially the same x-ray absorbance, they cannot be discriminated by a simple intensity threshold. Given the complex shapes in 3D and the presence of many connections between pores and cracks, shape discrimination is not successful either. Thus, this article proposes the use of Deep Convolutional Neural Networks (DCNN) to discriminate between these 2 classes of discontinuities. The well-known U-NET architecture was employed. The network was trained by manually outlining representative objects of the 2 classes in a few layers of the 3D image. After optimization of the training parameters, the network was applied to the full image, successfully discriminating between pores and cracks. The trained network was then applied to the images of different pellets with good results. However, some residual errors are present. These characteristics are analyzed and possible solutions are proposed