Effective segmentation of short fibers in glass fiber reinforced concrete’s X-ray images using deep learning technology

Abstract Fiber extraction and segmentation are critical to identifying the distribution of short fibers in fiber reinforced concrete. This work presents a new segmentation model based on the DeepLab V3 + network for automated probabilistic segmentation of fibers embedded in glass fiber reinforced concrete (GFRC) images obtained from micro-computed tomography (Micro-CT). A total of 2700 sliced images (715 × 715 pixels) are prepared by Micro-CT scanning and data augmentation, of which 2400 are for training with the remaining 300 for validation. It is shown that the proposed model can accurately and effectively segment the fibers in GFRC. The efficiency of the proposed segmentation model is accessed by three metrics of accuracy, intersection over union, and F1-score index, and they are reaching up to 99.3%, 80.4% and 67.2%, respectively. Moreover, the model can achieve a better segmentation than the traditional deep learning models. With good qualitative prediction, the proposed approach shows promise for predicting fiber segmentation based on unlabeled data obtained in the field. Finally, three-dimensional distribution of short fibers in GFRC samples with size of 2.99 × 2.99 × 2.99 mm3 is developed for reconstruction analysis.