Description of input data
Automatic defect detection in tomographic volumes using artificial intelligence approaches

The input data used in this article are tomographic volumes from various industrial partners, as well as from the Centre technique des industries de la fonderie (CTIF). The different manufacturers do not necessarily have the same equipment, and the X-ray tubes and detectors used are different, as are the magnifications. As a result, the volumes supplied have resolutions ranging from 150 µm to 450 µm per voxel (the actual spatial resolution being less good, since it is degraded by photon scattering in the scintillator). Imaging conditions also vary (high voltage and integration time) to suit different part thicknesses, so that image quality in terms of contrast resolution varies widely. It's interesting to have a database that's as rich as possible, and also as diverse as possible, as this enables the networks to generalize their decision-making. An example of an image from a tomographic...