Segmentation of computed tomography images and high-precision reconstruction of rubber composite structure based on deep learning

摘要

The internal fiber structure of composites determined via X-ray micro-computed tomography (mu-CT) is a key factor affecting their properties. However, the low resolution and blurred fiber-matrix interfaces of the mu-CT images obtained for large composite structures as well as the different grayscale distributions caused by the uneven thicknesses of their complex shapes considerably reduce the accuracy of the reconstructed model. Meanwhile, the relatively long scanning time of mu-CT limits the usability of this technique for in-situ loading tests. Therefore, in this study, an image segmentation method based on U-net convolution neural network is applied to the segmentation of mu-CT image of fabric rubber composite to realize the high-precision reconstruction of its internal fabric structure. This approach effectively minimizes the image noise caused by fast scanning and reduces the scanning time by 80%. The accuracy of the reconstructed model and the high efficiency of the developed method are verified through finite element analysis. Hence, the described mu-CT image processing method enables high-precision reconstruction of fiber-reinforced composites and provides a significant reduction in the mu-CT scanning time.