Detail-Enhanced High-Resolution Warping-Only Field Transforms for Particle Image Velocimetry

Abstract

Particle image velocimetry (PIV) is a technique for nonintrusive flow field measurement. Constrained by the cost volume component, existing deep-learning-based PIV methods face inherent bottlenecks in high-resolution scenarios. On the one hand, it is difficult to bear the memory load caused by high-resolution features and has to rely on multiple downsampling to use low-resolution features for modeling, resulting in the loss of small-scale flow details. On the other hand, the memory consumption when training and estimating high-resolution PIV images increases exponentially. Images with resolution above 2k <inline-formula> <tex-math notation="LaTeX">$\times 2$ </tex-math></inline-formula>k cannot be estimated directly, and the block estimation-result stitching strategy is adopted, which further destroys the physical consistency of the flow field during stitching. To address these issues, we propose the detail-enhanced high-resolution warping-only field transforms (DH-WAFT), a PIV-oriented optical flow estimation method optimized for high-resolution scenarios. DH-WAFT employs a feature extraction method tailored for small targets, which further eliminates redundant noise and reduces memory consumption. A multisource feature fusion module based on a residual network is designed to model the spatial correlation between large-scale flow fields and small-scale particles. A detail complement branch with dynamic weight is introduced to adaptively compensate the microdisplacement information in different resolution scenes. DH-WAFT is trained on publicly available synthetic PIV data and evaluated on both synthetic and experimental datasets, demonstrating superior accuracy and performance. On the <inline-formula> <tex-math notation="LaTeX">$256\times 256$ </tex-math></inline-formula> synthetic dataset, the average endpoint error (AEPE) of DH-WAFT is relatively reduced by 57.4 % compared with RAFT256-PIV and 47.3 % compared with Twin-PIVNet. In addition, this study also conducts spatial resolution experiments, which confirms that DH-WAFT is able to retain the fluid structure at a finer scale. More importantly, DH-WAFT shows excellent performance when dealing with high-resolution experimental data, covering <inline-formula> <tex-math notation="LaTeX">$512\times 512$ </tex-math></inline-formula>, 1k <inline-formula> <tex-math notation="LaTeX">$\times 1$ </tex-math></inline-formula>k, and 2k <inline-formula> <tex-math notation="LaTeX">$\times 2$ </tex-math></inline-formula>k levels. Finally, DH-WAFT carries out the necessary ablation experiment discussion on accuracy, memory consumption, and inference latency.