Although existing unsupervised particle image velocimetry (PIV) methods avoid the reliance on large-scale labeled flow data, they often suffer from low reconstruction accuracy. To address this, we propose a new unsupervised deep learning framework – UnLiteFlowNet with Multi-Scale Inception and Subpixel Upsampling (UnLiteFlowNet-MSI-SU). Built upon the classic LiteFlowNet, our method incorporates a multi-scale inception depthwise convolution module to enhance feature extraction and replaces traditional bilinear interpolation with a subpixel upsampling layer for finer reconstruction. The network is trained with an unsupervised loss function, including structural similarity loss to preserve image details and improve estimation accuracy. Experiments on synthetic datasets show that our method reduces the average endpoint error (AEE) by 11–21.4% for five typical flow scenarios compared to the baseline UnLiteFlowNet. Tests on real images from the third international PIV challenge further confirm its superior performance in reconstructing fine-scale flow structures.
Zhou et al. (Sun,) studied this question.