ConvNeXt-PIV: A modern convolutional framework for particle image velocimetry

Abstract

In recent years, deep learning–based Particle Image Velocimetry (PIV) models have made remarkable progress in estimating velocity fields of complex flows. Conventional CNN-based PIV networks, relying on small convolutional kernels and pooling layers, are limited in capturing global structures and large-scale motions. Transformer-based models mitigate this limitation but incur substantial computational and memory costs. To address these issues, we propose ConvNeXt-PIV, an encoder built on the ConvNeXt module, where depthwise convolution serves as the main feature extractor and large kernels expand the receptive field. By redesigning the feature hierarchy and raising the output to one quarter of the input resolution, the network better captures particle spacing and velocity-gradient details. Experimental results on Problem Class I demonstrate that ConvNeXt-PIV educes the average endpoint error (AEE) by 50.0%, 68.6%, and 37.6% on the Backstep, JHTDB Channel, and DNS Turbulence datasets, respectively, compared with the baseline method. Meanwhile, the number of model parameters is reduced to 1.7M parameters while maintaining a comparable computational cost, highlighting a favorable balance between accuracy and efficiency. These results verify the potential of modern convolutional architectures for fluid-motion estimation and provide a new design paradigm for lightweight and high-precision PIV modeling.