User-friendly end-to-end fiber identification for fiber-reinforced cementitious composites (FRCC) through deep learning

Authentic fiber distribution and orientation in fiber-reinforced cementitious composites (FRCC) is vital in study of cracking mechanism and tensile mechanics model. Existing analytical methods based on fluorescent micro-scopes or CT are complex, time-consuming and labor-intensive. This study presents a general and user-friendly end-to-end fiber identification method for FRCC through deep learning, which accomplished greatly accurate semantic segmentation of fibers (>100 mu m) with an ordinary SLR camera. The optimal model achieved MIoU of 98.93% and class accuracy of 99.998% for fiber. Specifically, the effect of pore, which is the possible interference to fiber identification in the cross-sections, was analyzed and the extracted multidimensional features were visualized and discussed. Furthermore, a case study on fiber distribution and orientation based on the proposed method was carried out. The results show that new advances are made in terms of the simplicity of modeling process, the accuracy and intuitiveness of results, and the user experience.

Automated summary

This study presents a general and user-friendly method for fiber identification in fiber-reinforced cementitious composites (FRCC) using deep learning, which achieves accurate semantic segmentation of fibers with an ordinary camera. The study also analyzes the effect of pore interference and visualizes and discusses the extracted features. A case study on fiber distribution and orientation using the proposed method shows new advances in modeling simplicity, result accuracy and intuitiveness, and user experience.