Individual fibre segmentation from time-resolved computed tomography images of fibre-reinforced composites using deep learning
Please login to view abstract download link
Fibre-reinforced composites (FRC) are widely used in engineering for their excellent mechanical properties, which mainly derive from their reinforcing fibres. However, during loading, individual fibre breaks develop into fibre break clusters of different sizes, which govern the tensile strength of the composite [1]. Fibre break models have therefore been developed to predict and study the fibre break and cluster development. Validating such models requires in-situ tracking of fibre breaks, which has only recently become feasible through time-resolved synchrotron radiation computed tomography (CT) [2]. However, time-resolved CT scans with acquisition times in the order of seconds suffer from significantly reduced image quality, such as higher noise levels or lower spatial resolutions, compared to static slow-acquisition CT scans. This study, therefore, employs a deep learning-based method, the U-Net-id algorithm [3], to segment fibres automatically in time-resolved tomographic images (see Fig. 1). We used the results to examine the geometric characteristics of the segmented fibres, including size, shape, and length, for slow- and fast-acquisition datasets. These datasets took 0.5 s and 500 s per scan to acquire, respectively. The results show that the U-Net-id method performs a high-fidelity segmentation on the slow-acquisition images, while the fast-acquisition CT images cannot fully capture the fibre geometry information. This result underlines the need for advanced image enhancement techniques such as super-resolution and image denoising.