Fibre breaks govern the strength of unidirectional composite materials under tension. The progressive development of fibre breaks is studied using in situ X-ray computed tomography, especially with synchrotron radiation. However, even with synchrotron radiation, the resolution of the time-resolved in situ images is not sufficient for a fully automated analysis of continuous mechanical deformations. We therefore investigate the possibility of increasing the quality of low-resolution in situ scans by means of super-resolution (SR) using 3D deep learning techniques, thus facilitating the subsequent fibre break identification. We trained generative neural networks (GAN) on datasets of high-(0.3 μm) and low-resolution (1.6 μm) statically acquired images. These networks were then applied to a low-resolution (1.1 μm) noisy image of a continuously loaded specimen. The statistical parameters of the fibre breaks used for the comparison are the number of individual breaks and the number of 2-plets and 3-plets per specimen volume. The fully automated process achieves an average accuracy of 82% of manually identified fibre breaks, while the semi-automated one reaches 92%. The developed approach allows the use of faster, low-resolution in situ tomography without losing the quality of the identified physical parameters.
Keywords: computed tomography; deep learning; fibre breaks; image quality; super-resolution.