Treatment of Motion Blur in High-Speed PIV using Deep Learning *

A technique based on deep learning that can reduce errors caused by motion blur for high-speed PIV is proposed. Synthetic images had the Monte-Carlo method (MCM) applied to them, in order to assess the error caused by blurry images of tracers. Longer particle streaks resulted in increased displacement errors (reaching 0.2 – 0.5 pixels) and outlier frequency (sometimes exceeding 8%). A new deblur filter was developed utilizing a generative adversarial network (GAN) with 1 million synthetic images. The filter, or generator, was verified using MCM data that was not learned. The outlier frequency was reduced to approximately 5%, and displacement error decreased below 0.25 pixels. This generator was applied to real blurry PIV images of a synthetic jet and significantly reduced the number of outlier vectors.