[论文解读] Unsupervised Learning of Particle Image Velocimetry

Mingrui Zhang, Matthew D. Piggott|arXiv (Cornell University)|Jul 28, 2020

Advanced Vision and Imaging参考文献 19被引用 7

一句话总结

本文提出了首个用于粒子图像测速（PIV）的无监督深度学习方法，通过光度、一致性和平滑性损失来估计密集速度场，无需真实标签数据。该方法在经典和有监督的PIV方法中表现具有竞争力，尤其在湍流和复杂流场等挑战性场景中表现更优。

ABSTRACT

Particle Image Velocimetry (PIV) is a classical flow estimation problem which is widely considered and utilised, especially as a diagnostic tool in experimental fluid dynamics and the remote sensing of environmental flows. Recently, the development of deep learning based methods has inspired new approaches to tackle the PIV problem. These supervised learning based methods are driven by large volumes of data with ground truth training information. However, it is difficult to collect reliable ground truth data in large-scale, real-world scenarios. Although synthetic datasets can be used as alternatives, the gap between the training set-ups and real-world scenarios limits applicability. We present here what we believe to be the first work which takes an unsupervised learning based approach to tackle PIV problems. The proposed approach is inspired by classic optical flow methods. Instead of using ground truth data, we make use of photometric loss between two consecutive image frames, consistency loss in bidirectional flow estimates and spatial smoothness loss to construct the total unsupervised loss function. The approach shows significant potential and advantages for fluid flow estimation. Results presented here demonstrate that our method outputs competitive results compared with classical PIV methods as well as supervised learning based methods for a broad PIV dataset, and even outperforms these existing approaches in some difficult flow cases. Codes and trained models are available at https://github.com/erizmr/UnLiteFlowNet-PIV.

研究动机与目标

• 为解决真实世界流体流动估计中可靠真实标签数据稀缺的挑战。
• 开发一种基于深度学习的PIV方法，无需标注训练数据，从而实现对复杂真实场景的泛化能力。
• 探究无监督学习是否能在多种流动条件下达到或超越有监督和经典PIV方法的性能。
• 评估在无监督框架中结合光度、一致性和平滑性损失对速度场估计的有效性。

提出的方法

• 该方法采用基于LiteFlowNet改进的轻量级流网络UnLiteFlowNet-PIV，通过无监督训练适配PIV任务。
• 总损失函数由三部分组成：光度损失以最小化变形图像之间的强度差异，一致性损失以强制双向流对称性，空间平滑性损失以促进速度场的平滑性。
• 光度损失通过参考图像与使用预测光流变形的第二张图像之间的L1范数计算。
• 一致性损失确保前向和后向流预测互为逆关系，从而提高鲁棒性。
• 平滑性损失通过惩罚预测流场中的大空间梯度来维持物理合理性。
• 模型仅使用PIV序列中的图像对进行端到端训练，无需任何真实速度标注。

实验结果

研究问题

• RQ1 无监督深度学习框架是否能在不依赖真实速度数据的情况下实现具有竞争力的PIV性能？
• RQ2 损失函数中的各个组成部分（光度、一致性、平滑性）如何分别对模型性能产生贡献？
• RQ3 在真实标签不可用的复杂或真实世界流动场景中，无监督方法是否比有监督方法具有更好的泛化能力？
• RQ4 在哪些类型的流动条件下，无监督方法能优于经典PIV和有监督深度学习基线方法？

主要发现

• 所提出的无监督方法在Back-Step、Cylinder、JHTDB、DNS和SQG数据集上的平均端点误差（AEE）分别为10.1、7.8、9.6、13.5和19.7像素每100像素，优于经典WIDIM和Horn-Schunck方法。
• 在具有挑战性的SQG和DNS数据集中，无监督方法的AEE低于所有有监督基线模型，包括最先进的有监督模型PIV-LiteFlowNet-en。
• 消融实验表明，若移除一致性或平滑性损失，性能将显著下降，证实了这两个组件在损失函数中的重要性。
• 完整的损失函数（$L_P + L_S + L_C$）取得最佳测试性能，Back-Step数据集上AEE为10.1，SQG数据集上为19.7，证明了联合损失设计的有效性。
• 该方法在复杂流动（如湍流和大位移）中泛化良好，而有监督模型因领域偏移问题而表现不佳。
• 无监督方法在所有数据集上均取得具有竞争力的结果，包括高雷诺数和复杂流动结构的数据集，且无需任何真实标签监督。