Unsupervised Particle Image Velocimetry Network Based on Improved LiteFlowNet3 with Embedded Cross-correlation and H-RAMi Attention Mechanism

Unsupervised particle image velocimetry (PIV) methods eliminate the need for extensive labeled datasets, yet their performance in complex flow regions, such as those with particle occlusion or out-of-boundary motion, remains limited. To address this issue, this study proposes an unsupervised PIV network: Unsupervised Enhanced LiteFlowNet3 with Cross-Correlation (UnLECNet-PIV). Built upon the LiteFlowNet3 architecture, the model incorporates structural refinements to its encoder-decoder design and integrates two core innovations: a Hierarchical Reciprocal Attention Mixer (H-RAMi) for multi-scale feature fusion and cross-correlation-derived feature priors to strengthen motion representation. Additionally, a boundary-aware loss function is introduced to refine flow estimation near image edges, complemented by a reflection padding strategy to preserve structural integrity at boundaries. Benchmark evaluations on synthetic PIV datasets demonstrate that UnLECNet-PIV reduces the average endpoint error (AEE) by approximately 25.9% compared to the baseline UnLiteFlowNet, while maintaining robustness and high spatial resolution under challenging conditions, including fine-scale vortical structures and Gaussian noise. Real experimental data further validate its physical consistency, underscoring the method’s practical effectiveness.