Space-time image velocity measurement method based on improved GhostNetV2

Abstract The space-time image velocimetry (STIV) method exhibits strong performance in non-contact measurements of river flow velocity and discharge. By identifying the main orientation angles of textures in space-time images and combining it with field data from hydrological stations, surface velocity in rivers can be measured. However, the generated space-time images are often affected by background noise and interfering textures due to video recording conditions or lighting variations, which impact the detection of texture orientation angles. To enhance the accuracy and stability of STIV, this paper proposes preprocessing techniques, including second-order differential frequency domain unsharp enhancement, to address noise and interference issues. Additionally, leveraging the correspondence in texture orientation between space-time images and two-dimensional autocorrelation images, a dataset of two-dimensional autocorrelation images with ninefold central magnification is created. An improved GhostNetV2 network is employed for angular classification of this dataset, resulting in increased classification accuracy that enhances the precision of river surface velocity estimation. Comparative experiments are conducted between existing methods and the proposed approach to validate its effectiveness. The experimental results show that the proposed method achieved relative errors in the average velocity measurements of 3.65% and 2.94% for two rivers, and relative errors in total discharge measurements of 3.73% and 2.08%, indicating greater consistency with the current meter measurements than other methods.