Automatic defect detection in tomographic volumes using artificial intelligence approaches

Like radiography, tomography is based on the differential attenuation of X-rays as a function of material density and chemical composition, but it exploits a large number of views taken from different angles by rotating the observed object. The different views enable us to determine the attenuation of each volume element, called a "voxel", and thus to reconstruct the object in three dimensions. Compared with radiography, which produces images of the volume projected onto the detector plane, tomography enables material to be examined in fictitious slices or sections. This avoids the numerous thickness variations and wall projections that are characteristic of radiography, and makes it much easier to recognize the nature of any discontinuities (or defects) present. This makes tomography an ideal tool for part development and expert appraisal. Thanks to the acceleration of computing resources, tomography is beginning to be used in production control.

The question of data interpretation then becomes a crucial issue. Indeed, to examine the entire volume, it is necessary to scroll through 2D virtual sections on the screen, or to use an algorithm enabling 3D representation of the object's volume, and interpreting the entire volume is very cumbersome to do manually. On a production line, automatic data processing is needed to detect discontinuities (such as missing material or inclusions). Such a task can be achieved with noise filtering and adaptive thresholding, but the performance achieved is the result of a compromise between detection of small defects, and detection of false alarms, due in particular to the fact that tomography is subject to reconstruction artifacts. The advent of convolutional neural networks, and the success achieved on natural images by deep networks, suggest that performance should be good in a non-destructive testing situation such as tomography.

This article proposes to demonstrate the usefulness of automatic defect detection methods in industrial tomography images using convolutional neural networks. The target application is aluminum casting, but other fields are possible, provided that a suitable database is defined.