POD-based background removal for Particle Image Velocimetry

POD background removal 

M.A. Mendeza, M. Raiolab, A. Masulloc, S. Discettib, A. Ianirob, R. Theunissenc, J.-M.Buchlina

a von Karman Institute for Fluid Dynamics, Waterloosesteenweg 72, Sint-Genesius-Rode, Belgium

b Aerospace Engineering Group, Universidad Carlos III de Madrid, Av. de la Universidad 30, Leganés, Spain

c Department of Aerospace Engineering, University of Bristol, University Walk, BS81TR, Bristol, UK


This website provides an implementation of the POD-based background removal algorithm described in the paper POD-based background removal for Particle Image Velocimetry. The method consists in approximating the background noise source and the PIV particle pattern with reduced order models (ROM) constructed from different portions of the video sequence’s POD spectra. Particles images and background noise are therefore distinguished according to a novel criterion: the higher degree of correlation of the background noise compared to the one of the particle pattern. Correlated background noise can be well approximated by a few of the first POD modes of the video, while the PIV particle pattern is equally distributed along the entire POD spectra. The proposed method is therefore a POD filter, which automatically identifies –and remove– the minimal number of modes representing the background noise.

POD Decomposition of PIV image recordings

Let a PIV image sequence be composed of n_t grayscale images Im(i,j)in mathbb{R}^{n_xtimes n_y} having a resolution of n_p=n_x n_y pixels. By reshaping each image into a column vector s_iin mathbb{R}^{n_ptimes 1}, it is possible to assemble the sequence into a snapshots matrix X:

(1)quad X={ s_1, s_2, dots s_{n_t}} in  mathbb{R}^{n_ptimes n_t}

The scope of low dimensional modeling of matrix X is to find the approximation tilde X in mathbb{R}^{n_ptimes n_t} of rank r<n_t minimizing the L_2 norm (||cdot||) of the error matrix E_r:

 (2)quad min(E_r)=min biggl (|| X-tilde X||_2 biggl ) ,.

The solution to this minimization problem, given by the Eckart-Young theorem, is the r truncated singular value decomposition of the original matrix:

 (3)quad tilde X=Phi_r ,  Sigma_r , { Psi^T_r} rightarrow tilde s_i=sum_{k=1}^{r}phi_k  sigma_k  psi_k^{i} , ,

with Phi_r=[phi_1,dotsphi_{r}] in mathbb{R}^{n_ptimes r} the orthonormal basis for the columns of X, Psi_r=[psi_1,dotspsi_{r}]in mathbb{R}^{n_ttimes r} the orthonormal basis for the rows of X, and Sigma_r=diag(sigma_1dotssigma_r)in mathbb{R}^{rtimes r} the diagonal matrix containing the norm of each contribution.

In low rank modeling for video analysis, the images forming the spatial basis phi are referred to as eigenbackgrounds. By definition, the phi_k are eigenvectors of the outer product matrix C=X X^T in mathbb{R}^{n_ptimes n_p} and the psi_k are eigenvectors of the inner product matrix K=X^T Xin mathbb{R}^{n_ttimes n_t}, while the singular values sigma_k are the square root of the corresponding eigenvalues lambda_k;

 (4a)quad C=X,X^T=bigl( Phi Sigma  Psi^Tbigr)bigl( Psi Sigma  Phi^T bigr)= Phi{Lambda} Phi^T
 (4b)quad K=X^T ,X=bigl(  Psi  Sigma  Phi^T bigr) bigl(  Phi  Sigma  Psi^Tbigr)=  Psi {Lambda}  Psi^T

The solutions to the eigenvalue problems expressed in eq.s (4a) and (4b) are the discrete versions of the Fredholm Equations, leading, respectively, to the definitions of standard POD (preferable when n_pll n_t) or the Snapshot POD (preferable when n_tll n_p). It should be noted that both definitions are common in the analysis of turbulent flows where instead of intensities, element entries of column vectors s_i refer to velocities.

Observing that {X} {Psi}_r={ Phi}_r  {Sigma}_r , eq. (3) can be also written as:

 (5)quad {tilde X}= Phi_{r} ,  Phi_{r}^T ,  {X} rightarrow tilde s_i=sum_{k=1}^{r}bigl( phi_k^T s_ibigr) phi_k , .

This form of the equation, with no emphasis on the temporal evolution of the modes, describes the decomposition as the projection of the data set (of rank n_t) into a lower dimensional space (of rank r<n_t) spanned by the orthonormal basis images  Phi_r=[phi_1,dotsphi_{r}]. This formulation is common in Principal Component Analysis where it is introduced in the framework of variance maximization or minimal error of the approximation matrix tilde X.

The POD image preprocessing proposed in this work considers a PIV sequence as the sum of an ideal sequence X_p (i.e. bright particle images superimposed onto a black background) and a background noise sequence X_b, each having their own singular value decomposition:

 (6)quad X={Phi} {Sigma} {Psi}^T=X_p+X_b={Phi}_p {Sigma}_p {Psi}_p^T +{Phi}_b {Sigma}_b {Psi}_b^T ,,,

with Phi_p=[phi_{p1},dots, phi_{p n_t}] and Phi_b=[phi_{b1},dots, phi_{b n_t}] the eigenbackgrounds of {X_p} and {X_b}. Typical background noise in PIV has a high degree of spatial and temporal correlation, resulting in multiple rows and columns of X_b being similar to each others. Therefore, the matrix X_b is close to be rank deficient and can be well captured by few (rll n_t) of its modes, such that

 (7)quad X_bapprox tilde{X}_b=sum_{k=1}^{r}phi_{bk}sigma_{bk}psi^T_{bk} , | , ,sigma_{bk}approx 0 , forall k>r ll n_t , ,,

with rank(tilde{X}_b)={rll n_t}. It is worth observing that, besides allowing for the background noise to be time dependent –contrary to simple levelization approaches– the proposed method also allows for the video sequence to be temporally unresolved –contrary to time filtering approaches–.

A temporally unresolved sequence can in fact be constructed from column permutation of a time-resolved sequence, and the SVD decomposition in eq. (3) –thus the approximation in eq. (7)– is invariant under column permutation of the decomposed matrix.

From eq.s (6-7), the proposed method consist in constructing an approximation of X_p and X_b using the POD modes of X. The method is based upon two assumptions, which are justified in the reference paper:

Assumption 1

For {k>r}, the contribution of the ideal PIV sequence X_p is equally distributed, such that sigma_{pk}approx sigma_{pk+1} , forall kin [r, n_t].

Assumption 2

For {k>r}, the decomposition of the video X is aligned with that of the ideal PIV sequence X_p, such that sigma_{k}approx sigma_{pk} , forall kin [r, n_t].

Proposed Algorithm

Since rll n_t and sigma_{pk}approx sigma_{pk+1} (Assumption 1), it is possible to approximate the ideal PIV video sequence X_p underlying the video sequence X (eq. (6)) filtering out its first r POD modes:

 X_p= sum_{k=1}^{n_t} phi_{pk} sigma_{pk} psi_{pk}^Tapproxtilde{X}_p=sum_{k=r+1}^{n_t} phi_{pk} sigma_{pk} psi_{pk}^T ,, .

Moreover, since sigma_{k}approx sigma_{pk} , forall kin [r, n_t] (Assumption 2), it is reasonable to expect the decomposition of X to be aligned with that of X_p for k>r. Therefore, using eq. (5} yields:

 tilde{X}_papprox tilde{X}=sum_{k=r+1}^{n_t} phi_{k} sigma_{k} psi_{k}^T =tilde {Phi} tilde {Phi}^T X ,

where tilde {Phi}=[phi_{r+1},dots phi_{n_t}] is the basis for the reduced order model of X_p. In addition to the equality of singular values, the modes approximating the PIV pattern should have a temporal psi_k, for {k>r}, orthonormal to psi_{p1}=underline{1}, i.e. langle psi_{k},underline{1} rangle=sum_{j=1}^{n_t} psi^j_{k}=0, as discussed in the paper. These two constraints are used to identify the [r+1,n_t] POD modes approximating the PIV pattern, to be retained in the preprocessing. Then, the method consists in constructing the reduced basis onto which project the set of images. If significant light variations appear between two frames a and b, the method should be applied independently on the two series of camera exposures. The pseudo-code of the proposed method is summarized in the following algorithm, where the tolerances in line 7 are set as varepsilon_1=0.01sigma_{pk}=0.01sqrt{n_p}sigma_{sp} and varepsilon_2=0.01:

  1. Reshape Images Imin mathbb{R}^{n_xtimes n_y} in s_iin mathbb{R}^{n_ptimes 1}}$

  2. Assemble Matrix Xin mathbb{R}^{n_ptimes n_t}}$

  3. Compute K=X^T X}$

  4. Diagonalize K=Psi , Sigma^2 , Psi^T }$

  5. Compute Phi=X , Psi , Sigma^{-1}}$

  6. Find r: sigma_{k+1}-sigma_k<varepsilon_1 , & , langle underline{1}, psi_{pk}rangle < varepsilon_2 ,, forall k>r}$label{line}

  7. Construct tilde{Phi}=[phi_{r+1},dots phi_{n_t}]}$

  8. Compute tilde{X}=tilde{Phi}tilde{Phi}^T X with tilde{X}=[tilde{x_1},dots tilde{x_{n_t}}]}$

  9. Reshape tilde{s}_iin mathbb{R}^{n_ptimes 1} back to tilde{Im}in mathbb{R}^{n_xtimes n_y}}$

Analysis of statistical convergence on synthetic PIV images 

Analysis of statistical convergence on synthetic PIV images with source density N_S=0.02 and N_S=0.9.