Multi-stage image processing approach for 3D-CT based in- or at-line inspection of complex parts

In the recent years, the use of 3D computed tomography (3D-CT) as an industrial inspection method has grown significantly. While the application of 3D-CT as a non-destructive inspection and metrology method has been established in laboratories for decades, it took a long time for 3D-CT to actually find its way into the production floor. One reason is that standard analysis software quite often fails under production floor conditions. As typical, cycle times of production are short and lot sizes in production may be huge, and the harsh imaging conditions differ significantly from those in a laboratory environment. That is why simple adaptation and automation of software developed for manual use in laboratories is not always sufficient. In order to catch up with cycle times in production, be it in-line, at-line or during large batch volume sample testing, 3D-CT scans have to be performed very fast. To achieve this, the number of acquired projections per scan and the integration time per projection must be heavily reduced. This results in a significantly reduced signal-to-noise ratio and loss of contrast resolution which makes it especially hard to detect low contrast discontinuities and anomalies such as porosities or low contrast inclusions. But as well small inclusions with high contrast may be hidden by the high noise level. Moreover, it is of course not possible to perform more than one scan per component, as one would perhaps do in a laboratory to achieve optimal results. There is simply no time to perform multiple scans in different orientations to optimize X-ray penetration on a dedicated part of a complex shaped object or to avoid edge artefacts. Dual-energy scans for metal artefact reduction normally are not taken into account as well. All this leads to the situation that 3D-CT scans out of production are noisy and are often contaminated by different forms of artefacts, which becomes a challenge for any inspection software. Especially classical, mainly analytical image processing, which is as a standard used in a laboratory quite often, does not deliver acceptable performance on such scans. Even when it is possible to use classical image processing, different settings require extensive fine tuning and adjustments to accommodate different regions of interest. In any case, the operator has to deal with a large amount of false positive detections which require a lot of manual rework. This situation made it at least difficult to use 3D-CT in production and has delayed its widespread usage. The situation began to change a couple of years ago with the emergence of AI-based analysis methods that complemented classical image processing. For example, machine learning based image processing approaches or convolutional neural networks (CNNs) are better at handling noisy data and data containing artefacts compared to classical image processing. Quite often their defect sensitivity is higher, and the amount of false positive detections is ...