On Uncertainty Prediction for Deep-Learning-based Particle Image Velocimetry

TL;DR

This paper evaluates three methods for quantifying uncertainty in deep learning-based Particle Image Velocimetry, demonstrating that the Uncertainty neural network (UNN) outperforms others in accuracy and reliability across multiple datasets.

Contribution

It introduces and compares three uncertainty quantification methods for deep learning PIV, highlighting the effectiveness of UNN for reliable uncertainty estimates.

Findings

UNN provides the most accurate uncertainty estimates.

All methods perform well under mild perturbations.

UNN shows strong potential for practical PIV applications.

Abstract

Particle Image Velocimetry (PIV) is a widely used technique for flow measurement that traditionally relies on cross-correlation to track the displacement. Recent advances in deep learning-based methods have significantly improved the accuracy and efficiency of PIV measurements. However, despite its importance, reliable uncertainty quantification for deep learning-based PIV remains a critical and largely overlooked challenge. This paper explores three methods for quantifying uncertainty in deep learning-based PIV: the Uncertainty neural network (UNN), Multiple models (MM), and Multiple transforms (MT). We evaluate the three methods across multiple datasets. The results show that all three methods perform well under mild perturbations. Among the three evaluation metrics, the UNN method consistently achieves the best performance, providing accurate uncertainty estimates and demonstrating strong potential for uncertainty quantification in deep learning-based PIV. This study provides a comprehensive framework for uncertainty quantification in PIV, offering insights for future research and practical implementation.