PIV-FlowDiffuser: Transfer-Learning-Based Denoising Diffusion Models for Particle Image Velocimetry.

Area of Science:

• Fluid dynamics
• Computational fluid dynamics
• Image analysis

Background:

• Deep learning enhances particle image velocimetry (PIV) computational speed and resolution.
• Domain gaps between synthetic training data and real-world images degrade PIV model performance, introducing noise.
• Residual patterns in vector fields are common issues in deep learning-based PIV estimators.

Purpose of the Study:

• To introduce a novel method, PIV-FlowDiffuser, for reducing noise in PIV analysis using denoising diffusion models.
• To improve the accuracy and generalization capabilities of PIV algorithms on practical particle images.

Main Methods:

• Employed a denoising diffusion model (FlowDiffuser) for iterative noise reduction in PIV.
• Trained a data-hungry iterative denoising diffusion model using a transfer learning strategy.
• Pre-trained the FlowDiffuser model on diverse optical flow datasets (e.g., Sintel, KITTI) and fine-tuned it on synthetic PIV datasets, upsampling images by 2x.

Main Results:

• PIV-FlowDiffuser effectively suppressed noise patterns in visualized vector fields.
• Achieved a 59.4% reduction in average endpoint error (AEE) compared to the RAFT256-PIV baseline on Cai's dataset.
• Demonstrated enhanced generalization performance on unseen particle images due to transfer learning.

Conclusions:

• Denoising diffusion models, particularly when enhanced with transfer learning, offer a powerful approach for improving PIV accuracy.
• The PIV-FlowDiffuser method effectively addresses domain gap issues and noise in practical PIV applications.
• Highlights the potential of transfer-learning-based denoising diffusion models for advancing PIV analysis.