E-Thesis 326 views
Computer Vision-based Microstructure Reconstruction of Heterogeneous / XIANGYUN GE
Swansea University Author: XIANGYUN GE
DOI (Published version): 10.23889/SUThesis.69222
Abstract
This thesis presents a novel approach to heterogeneous material microstructure image re-construction using AI-generated content (AIGC) methods based on computer vision (CV). By focusing on the unique challenges posed by heterogeneous material structures, this work optimizes network architectures to...
| Published: |
Swansea University, Wales, UK
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Neto, E. D. S., and Li, C. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa69222 |
| Abstract: |
This thesis presents a novel approach to heterogeneous material microstructure image re-construction using AI-generated content (AIGC) methods based on computer vision (CV). By focusing on the unique challenges posed by heterogeneous material structures, this work optimizes network architectures to enhance both the efficiency and accuracy of image generation. The research can be divided into three major contributions.First, the Statistically-Informed Neural Networks (SINN) method, originally designed for binary-phase porous media, is extended to multi-phase heterogeneous materials. This ex- tension broadens the applicability of the SINN approach, enabling more complex material reconstructions.Second, the explicit descriptor-based optimization techniques from traditional statistical reconstruction algorithms are incorporated into CV-based AI microstructure reconstruction methods. This integration ensures that the generated images not only match geometric pat- terns but also adhere closely to quantitative material descriptors, improving interpretability and precision.Third, improvements are made to existing CV-based generative models by tailoring them specifically to the task of heterogeneous material microstructure reconstruction. These improvements simplify the model structure and reduce the number of parameters while maintaining high accuracy in image generation.Ultimately, the combined methods lead to the first-ever conditional generative model capable of generating 3D images from 2D slice descriptors. The proposed model is validated throughcomparison with the state-of-the-art SliceGAN, demonstrating superior accuracy and efficiency through explicit optimization. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information |
| Keywords: |
Image reconstruction, heterogeneous material, microstructure reconstruction |
| College: |
Faculty of Science and Engineering |
| Funders: |
China Scholarship Council |