Sensors & Transducers



Vol. 249, Issue 2, February 2021, pp. 25-35





1, * Bernhard LEHNER, 2 Thomas GALLIEN, 1, 4 Péter KOVÁCS,
5 Gregor THUMMERER, 5 Günther MAYR, 6 Peter BURGHOLZER
​and 1, 3 Mario HUEMER



1 Silicon Austria Labs (SAL), JKU LIT SAL eSPML Lab, Altenbergerstr. 69,
4040, Austria

2 Silicon Austria Labs (SAL), Inffeldgasse. 33, 8010, Austria

3 Johannes Kepler University Linz, Institute of Signal Processing, JKU LIT SAL eSPML Lab, Altenbergerstr. 69, 4040, Austria

4 Department of Numerical Analysis, Eötvös Loránd University, 1117, Hungary

5 Josef Ressel Centre of Thermal NDE of Composites, University of Applied Sciences Upper Austria, Roseggerstraße 15, 4600, Austria

6 Research Center for Non Destructive Testing, Altenberger Straße 69, 4040, Austria

1 Tel.: +43 5 0317

E-mail: bernhard.lehner@silicon-austria.com



Received: 1 October 2020 /Accepted: 15 January 2021 /Published: 28 February 2021





Abstract: Thermographic imaging is a contactless and nondestructive way to detect defects inside the specimen. The current state-of-the-art approach combines model- and deep learning-based reconstructions in a hybrid fashion. The recently developed virtual wave concept (VWC) provides a framework to develop such hybrid solutions, and allows to utilize physical priors, such as non-negativity and/or sparsity. In combination with the superiority of deep learning approaches over hand-crafted features and heuristics, improved reconstruction accuracy compared to previous methods was achieved. However, the reconstruction results still deteriorate under low SNR conditions caused by defects located deeper underneath the surface. Therefore, it would be useful to have an uncertainty estimate that reflects the reliability of the reconstructions. This would enable the automatic identification of results that are likely to be inaccurate and require a closer inspection. In this paper, we propose two computationally very cheap methods to estimate the uncertainty of thermographic imaging results. In order to show the generalization capability of our approach, we thoroughly evaluate it under different conditions and even with different deep model architectures.


Keywords: Thermal tomography, Virtual wave, Nondestructive testing, Regularization, Sparsity, Deep learning.

​___________________________________________________________________