An improved YOLOv8 for fiber bundle segmentation in X-ray computed tomography images of 2.5D composites to build the finite element model

Abstract

It is necessary to segment fiber bundles in the reconstruction of the mesoscopic model of ceramic matrix composites using XCT images. Existing methods have great subjectivity, poor recognition accuracy, and heavy workload. To solve this problem, an improved lightweight YOLOv8 was proposed, which is a deep learning approach. By adding Slim-neck and VanillaNet, the complexity of the model was greatly reduced. Additionally, by replacing the loss function of the model with the Wise-IoU loss function, the ability of feature extraction of the model was improved. The effectiveness of the improved YOLOv8 in fiber bundle identification was demonstrated. Finally, a mesoscopic model was reconstructed by XCT images where fiber bundles were segmented by using the improved YOLOv8. The linear elastic modulus of the material was predicted and the error was found to be small, indicating that the improved YOLOv8 can effectively segment fiber bundles and thus reconstruct a high-precision mesoscopic model.