PIV anisotropic denoising using uncertainty quantification

Recently, progress has been made to reliably compute uncertainty estimates for each velocity vector in planar flow fields measured with 2D-or stereo-PIV. This information can be used for a post-processing denoising scheme to reduce errors by a spatial averaging scheme preserving true flow fluctuations. Starting with a 5 × 5 vector kernel, a second-order 2D-polynomial function is fitted to the flow field. Vectors just outside will be included in the filter kernel if they lie within the uncertainty band around the fitted function. Repeating this procedure, vectors are added in all directions until the true flow field can no longer be approximated by the second-order polynomial function. The center vector is then replaced by the value of the fitted function. The final shape and size of the filter kernel automatically adjusts to local flow gradients in an optimal way preserving true velocity fluctuations above the noise level. This anisotropic denoising scheme is validated first on synthetic vector fields varying spatial wavelengths of the flow field and noise levels relative to the fluctuation amplitude. For wavelengths larger than 5–7 times the spatial resolution, a noise reduction factor of 2–4 is achieved significantly increasing the velocity dynamic range. For large noise levels above 50% of the flow fluctuation, the denoising scheme can no longer distinguish between true flow fluctuations and noise. Finally, it is shown that the procedure performs well for typical experimental PIV vector fields. It provides an effective alternative to more complicated adaptive PIV algorithms optimizing interrogation window sizes and shapes based on seeding density, local flow gradients, and other criteria.