LIU, XINGYU (2025) Enhanced Flow Visualization Using Image Processing and Deep Learning Techniques. Doctoral thesis, Durham University.
![]()
| PDF - Accepted Version 7Mb |
Abstract
This thesis explored the application of machine learning techniques to enhance the efficiency and accuracy of surface flow visualization (SFV). The SFV technique has been widely used in fluid dynamics research to provide qualitative information. In order to extract some quantitative information from SFV images, specific algorithms must be developed. While machine learning algorithms are a type of algorithm that automatically analyses data to obtain patterns and uses these patterns to make prediction.
The core innovation of this thesis lies in the use of Convolutional Neural Networks (CNNs) to automate streamline detection, demonstrating superior reliability and accuracy compared to traditional methods like Sobel edge detection. Building on this, the thesis proposes a predictive neural network model capable of estimating flow fields from SFV images. To train this model, a comprehensive dataset was constructed using both experimental data and synthetically generated images, significantly improving the model's robustness and generalization ability.
The Global Luminescent Oil-Film (GLOF) method has been conducted to determine if the surface friction fields could be extracted from the videos. The results from this thesis indicate that this is not the case. while GLOF has promising applications, substantial refinement is required to achieve reliable results in complex flow cases.
Therefore, the SFV images was labelled manually using chaincode method. And Generative Adversarial Networks (GANs) were applied to generate synthetic flow field images, which supplemented the experimental data and improved model training. Additionally, a simplified approach combined synthetic and experimental data to train predictive models like U-Net, improving the accuracy of flow field estimation.
this work delivers a practical framework that enables researchers to input a single SFV image and obtain a preliminary prediction of streamlines and flow fields. Another useful contribution is the creation of a unique dataset, hosted on GitHub, which combines experimental and synthetic data, enabling the training of flow visualization algorithms.
Item Type: | Thesis (Doctoral) |
---|---|
Award: | Doctor of Philosophy |
Keywords: | surface flow visualization, machine learning, GANs |
Faculty and Department: | Faculty of Science > Engineering, Department of |
Thesis Date: | 2025 |
Copyright: | Copyright of this thesis is held by the author |
Deposited On: | 11 Jun 2025 10:59 |