A lightweight neural network designed for fluid velocimetry

Abstract

We devise a novel Lightweight Image Matching Architecture (LIMA), which is designed and optimized for Particle Image Velocimetry (PIV). LIMA is a convolutional neural network (CNN) that performs symmetric image matching and employs an iterative residual refinement strategy, which allows us to optimize the total number of refinement steps to balance accuracy and computational efficiency. The network is trained on kinematic datasets with a loss function that penalizes larger gradients. We consider a six- (LIMA-6) and a four-level (LIMA-4) version of the network and demonstrate that they are considerably leaner and faster than a state-of-the-art network designed for optical flow (PWCIRR). LIMA-6 reconstructs the velocity field from synthetic and experimental PIV images with an accuracy comparable or superior both to PWCIRR as well as to state-of-the-art cross-correlation methods (i.e., a commercial implementation of WIDIM). Although less accurate, LIMA-4 allows a significant reduction of the computational costs with respect to any other method considered. All CNNs prove more robust than WIDIM with respect to particle loss and allow effective error reduction by increasing the particle seeding density. Thanks to reduced computational cost and memory requirement, we envision the deployment of LIMA on low-cost devices to provide affordable, real-time inference of the flow field during PIV measurements.

Autonomous Target Recognition for Low-Altitude UAVs Based On Slimmable Neural Networks

“…2 of 18 computational capabilities on unmanned aerial vehicle (UAV) platforms, it is crucial to enrich methods for aerial target recognition and detection by introducing Slimmable neural networks and incorporating the Squeeze and Excitation (SE) and Swin-Transformer mechanisms. These enhancements strengthen the model's adaptability to changes in perspective and scale, enabling it to possess stronger recognition capabilities [23]. Finally, comparative experiments using the publicly available VisDrone19 dataset have confirmed that both the SE-YOLOv5s and ST-YOLOv5s models outperform the YOLOv5s model in terms of performance [24].…”