Twins-PIVNet: Spatial attention-based deep learning framework for particle image velocimetry using Vision Transformer

Particle Image Velocimetry (PIV) for flow visualization has advanced with the integration of deep learning algorithms. These methods enable end-to-end processing, extracting dense flow fields directly from raw particle images. However, conventional deep learning-based PIV models, which predominantly rely on convolutional architectures, are limited in their ability to utilize contextual information and capture dependencies between pixels across sequential images, impacting the prediction accuracy. We introduce Twins-PIVNet, a deep learning framework for PIV optical flow estimation that leverages a spatial attention-based vision transformer architecture. Its self-attention mechanism captures multi-scale features of particle motion, significantly improving the dense flow field estimation. Trained on synthetic PIV datasets covering a wide range of flow conditions, Twins-PIVNet has been evaluated on both synthetic and experimental datasets, demonstrating superior accuracy and performance. In comparative studies, Twins-PIVNet outperforms existing optical flow and conventional methods, achieving accuracy improvements of 51% for backstep flow, 42% for DNS-turbulence, and 33% for surface quasi-geostrophic flow. Additionally, it also exhibits strong generalization on experimental PIV data, demonstrating robustness in handling real-world PIV uncertainties. Despite its attention mechanism, Twins-PIVNet maintains faster inference and training times compared to other PIV models, offering an optimal balance between complexity, efficiency, and performance.