A Lightweight Deep Optical Flow Network With Soft Threshold for Particle Image Velocimetry

Particle image velocimetry (PIV) is a key technology in experimental fluid dynamics, which enables estimating global physical displacement fields from consecutive image data for complex industrial problems. Despite the advances in the past decades, classical PIV estimators rely on artificially designed parameters and postprocessing procedures, which may be troublesome. Although deep learning (DL) models have provided an automatic framework to achieve similar performance, they are transitionally dependent on the datasets and difficult to be extended in challenging experimental scenarios. To this end, a generalized lightweight optical flow (OF) network with embedded soft thresholds is proposed in this article. The lightweight model is optimized in a teacher-student framework, where we incorporate the teacher’s response and features into the distillation loss to mitigate the performance degradation caused by model compression. In addition, the attention mechanism is deployed to construct channel soft threshold networks, thus the features without useful information can be filtered by the network. The thresholds are considered as trainable parameters and learned during the training process. Benefiting from the threshold segmentation, the dilemma of sensitive to unfavorable factors such as noise, obstacles, and illuminance imbalance in PIV estimators is alleviated. Comprehensive evaluations, including tests on synthetic benchmark datasets and challenging experimental datasets with complex physical scenarios or poor image quality, demonstrate that the proposed model exhibits state-of-the-art accuracy and efficiency compared to previous DL-based OF models and classical PIV approaches. All the codes are available at https://github.com/wuwuwuas/LRAFT-ST.git