A Deep Learning-Based Approach To Increase Efficiency in the Acquisition of Ultrasonic Non-Destructive Testing Datasets

Nick Luiken, Matteo Ravasi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2023, pp. 3094-3102

Abstract


Ultrasonic phased array systems traditionally acquire data in a sequential fashion. Although using different excitation delays for each pulsing element can be used to steer the emitted wavefield into e.g., plane waves or focused beams, the overall frame rate of the system is dominated by the choice of the firing time between two consecutive experiments. Inspired from a technology in reflection seismology, we propose the use of simultaneous shooting to increase the flexibility of acquiring ultrasonic data in non-destructive testing (NDT) applications. Simultaneous shooting is an acquisition setup whereby separate transmit sequences are performed simultaneously at a reduced time interval leading to entangled data that may yield artifacts in subsequent imaging products. The data can be untangled by a process called deblending, which is a heavily underdetermined linear inverse problem. We solve the deblending problem using the recently introduced SSDeblend algorithm. This algorithm combines the physics of simultaneous shooting with a powerful self-supervised denoiser specifically tailored to remove the so-called blending noise. We conduct an experiment on an openly available Full Matrix Capture dataset and show that one can speed up the acquisition by at least a factor of 2 with little loss of quality on the resulting final image.

Related Material


[pdf]
[bibtex]
@InProceedings{Luiken_2023_CVPR, author = {Luiken, Nick and Ravasi, Matteo}, title = {A Deep Learning-Based Approach To Increase Efficiency in the Acquisition of Ultrasonic Non-Destructive Testing Datasets}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2023}, pages = {3094-3102} }