Improving the performance of a lightweight convolutional neural network for particle image velocimetry through hyper-parameter and padding optimization

Over the last decade, convolutional neural networks (CNNs) have gained popularity for applications to optical flow estimation. Recently, we proposed a novel lightweight image matching architecture (LIMA), a CNN specifically designed to process recordings from particle image velocimetry (PIV) and infer displacement fields in a fluid flow. LIMA demonstrated higher accuracy and lower training cost than select state-of-the-art CNNs in benchmark flow scenarios. However, further improvements are required to handle increasingly challenging PIV conditions, such as higher noise levels, complex flow geometries, and small vortex structures, which are critical aspects in most PIV experiments relevant in science and engineering. To enhance LIMA's performance, in this work, we examine the network's hyperparameters unique to PIV applications. We identify appropriate image padding that suppresses spurious artifacts propagating from image boundaries into the reconstructed displacement field during image convolution. We also discuss the correlation search range and show that it can be decreased to reduce the computational burden of LIMA without degrading inference accuracy. We demonstrate that, with the appropriate selection of the two parameters, LIMA offers enhanced capability to accurately reconstruct displacement fields in challenging and realistic PIV flow scenarios. We compare LIMA's performance with correlation-based PIV processing techniques and, for the first time, with another CNN proposed recently for PIV. These findings support the use of LIMA in advanced engineering applications, particularly ones that require both computational efficiency and high spatial resolution.