Phase segmentation of uncured prepreg X-Ray CT micrographs

This paper explores methods to investigate voids and dry fibre areas in pre-impregnated aligned carbon fibre reinforced epoxy lay-ups. A deep learning segmentation approach was compared to conventional thresholding techniques to characterise the interlaminar voids (entrapped air) and dry areas (unsaturated fibre bed) phases obtained by micro-CT scanning of samples from uncured laminates. The performance of both approaches was quantitatively assessed in three regions of interest having different levels of porosity, ranging from a low 1% to a high 25%. Deep learning consistently outperformed thresholding in the segmentation of both interlaminar voidage and dry areas. Furthermore, deep learning improved the ability to detect small voids and was able to accurately segment voids in volumes with less than 2% voidage, whereas thresholding techniques fail in this task. Finally, the application of deep learning to the segmentation of dry areas in micro-CT scans provided sharper results than thresholding, without needing filtering.

Automated summary

This study compared the performance of deep learning segmentation approach with conventional thresholding techniques in investigating voids and dry fiber areas in pre-impregnated aligned carbon fiber reinforced epoxy lay-ups. It was found that deep learning consistently outperformed thresholding in segmenting interlaminar voidage and dry areas, especially in detecting small voids and segments with low porosity levels.