Beyond incoherence and beyond sparsity: compressed sensing in the real world

Size: px

Start display at page:

Download "Beyond incoherence and beyond sparsity: compressed sensing in the real world"

Barnard Dickerson
5 years ago
Views:

1 Beyond incoherence and beyond sparsity: compressed sensing in the real world Clarice Poon 1st November 2013

2 University of Cambridge, UK Applied Functional and Harmonic Analysis Group Head of Group Anders Hansen Research Associates Jonathan Ben-Artzi, Bogdan Roman PhD Students Alex Bastounis, Milana Gataric, Alex Jones, Clarice Poon Interests within sampling theory Computations in infinite dimensions, Inverse Problems, Nonuniform sampling Compressed sensing, including applications to medical imaging (MRI and CT), seismic tomography.

3 University of Cambridge, UK Applied Functional and Harmonic Analysis Group Head of Group Anders Hansen Research Associates Jonathan Ben-Artzi, Bogdan Roman PhD Students Alex Bastounis, Milana Gataric, Alex Jones, Clarice Poon Interests within sampling theory Computations in infinite dimensions, Inverse Problems, Nonuniform sampling Compressed sensing, including applications to medical imaging (MRI and CT), seismic tomography. Related Group: Cambridge Image Analysis - lead by Carola Schönlieb. Image and video processing via PDE and variational based methods

4 Outline Introduction A new theory for compressed sensing (Joint work with Ben Adcock, Anders Hansen and Bogdan Roman) Compressed sensing with non-orthonormal systems

5 Outline Introduction A new theory for compressed sensing (Joint work with Ben Adcock, Anders Hansen and Bogdan Roman) Compressed sensing with non-orthonormal systems

6 A reconstruction problem in C N We are given two orthonormal bases Ψ = { ψ j C N : j = 1,..., N }, Φ = { φ j C N : j = 1,..., N } We want to recover x C N where the underlying signal f is f = N j=1 x jφ j. We are given access to samples ˆf = ( f, ψ j ) N j=1.

7 A reconstruction problem in C N We are given two orthonormal bases Ψ = { ψ j C N : j = 1,..., N }, Φ = { φ j C N : j = 1,..., N } We want to recover x C N where the underlying signal f is f = N j=1 x jφ j. We are given access to samples ˆf = ( f, ψ j ) N j=1. We can construct a measurement matrix U C N N with entries u ij = φ j, ψ i. Ux = ˆf

8 A reconstruction problem in C N We are given two orthonormal bases Ψ = { ψ j C N : j = 1,..., N }, Φ = { φ j C N : j = 1,..., N } We want to recover x C N where the underlying signal f is f = N j=1 x jφ j. We are given access to samples ˆf = ( f, ψ j ) N j=1. We can construct a measurement matrix U C N N with entries u ij = φ j, ψ i. P Ω Ux = P Ωˆf where P Ω is the projection matrix onto the index set Ω {1,..., N}.

9 Current theory IF we have 1. Sparsity 2. Incoherence {j : x j 0} = s << N, µ(u) := max φ j, ψ k 2 = O ( N 1) j,k=1,...,n Then by choosing s log(n) log(ɛ 1 + 1) samples uniformly at random, x is recovered exactly with probability exceeding 1 ɛ by solving min β 1 subject to P Ω Uβ = P Ωˆf. β C N

10 Current theory IF we have 1. Sparsity 2. Incoherence {j : x j 0} = s << N, µ(u) := max φ j, ψ k 2 = O ( N 1) j,k=1,...,n Then by choosing s log(n) log(ɛ 1 + 1) samples uniformly at random, x is recovered exactly with probability exceeding 1 ɛ by solving min β 1 subject to P Ω Uβ = P Ωˆf. β C N In MRI, we are interested in Fourier sampling with wavelet sparsity. But, the coherence between Fourier and any wavelet basis is µ = 1. There are many important applications which are highly coherent. e.g. X-ray tomography, electron microscopy, reflection seismology.

11 Uniform random sampling min β C N β 1 subject to P Ω U df V 1 dw β = P Ωˆf. 5% subsampling map Original (1024x1024) Test phantom constructed by Guerquin-Kern, Lejeune, Pruessmann, Unser, 12

12 Uniform random sampling 5% subsampling map Reconstruction (1024x1024) (1024x1024)

Variable density sampling 5% subsampling map Reconstruction (1024x1024) (1024x1024) Empirical solution (Lustig (2008), Candès (2011), Vandergheynst

13 Variable density sampling 5% subsampling map Reconstruction (1024x1024) (1024x1024) Empirical solution (Lustig (2008), Candès (2011), Vandergheynst (2011)): take more samples at lower Fourier frequencies and less samples at higher Fourier frequencies. Q: How can we theoretically justify this approach?

14 Outline Introduction A new theory for compressed sensing (Joint work with Ben Adcock, Anders Hansen and Bogdan Roman) Compressed sensing with non-orthonormal systems

15 Multi-level sampling Divide the samples into r levels and let N = (N 1,..., N r ) with 1 N 1 <... < N r = N, m = (m 1,..., m r ) with m k N k N k 1, N 0 = 0, Ω k {N k 1 + 1,..., N k } is chosen uniformly at random, Ω k = m k, We refer to the set as an (N, m)-sampling scheme. Ω N,m = Ω 1... Ω r

16 To understand the use of multi-level random sampling instead of uniform random sampling, we replace the standard ingredients of compressed sensing with more realistic assumptions: Sparsity Asymptotic sparsity Incoherence Asymptotic incoherence

Asymptotic sparsity Divide the reconstruction coefficients into levels and consider sparsity at each level: For r N, let M = (M 1,..., M r ) with 1 M 1 <... < M r = N, s = (s 1,.

17 Asymptotic sparsity Divide the reconstruction coefficients into levels and consider sparsity at each level: For r N, let M = (M 1,..., M r ) with 1 M 1 <... < M r = N, s = (s 1,..., s r ) with s k M k M k 1, M 0 = 0, We say that x C N is (M, s)-sparse if for each k = 1,..., r, supp(x) {M k 1 + 1,..., M k } s k Left: test image, Right: Percentage of wavelet coefficients at each scale which are greater than 10 3 in magnitude.

18 Asymptotic incoherence We say that U C N N is asymptotically incoherent if for large N, µ(p K U), µ(up K ) = O ( K 1) where P K is the projection matrix onto the index set {1,..., K}. For any wavelet reconstruction basis with Fourier samples µ(u) = 1, µ(p K U), µ(up K ) = O ( K 1). Implication of asymptotic incoherence: Sample more at low Fourier frequencies where the local coherence is high, and subsample at higher Fourier frequencies.

19 Asymptotic incoherence We say that U C N N is asymptotically incoherent if for large N, µ(p K U), µ(up K ) = O ( K 1) where P K is the projection matrix onto the index set {1,..., K}. For any wavelet reconstruction basis with Fourier samples µ(u) = 1, µ(p K U), µ(up K ) = O ( K 1). Implication of asymptotic incoherence: Sample more at low Fourier frequencies where the local coherence is high, and subsample at higher Fourier frequencies. Notation: Given vectors M = (M 1,..., M r ) and N = (N 1,..., N r ) with 0 = M 0 < M 1 < < M r = N and 0 = N 0 < N 1 <... < N r = N, let µ N,M (U)[k, l] := µ(p N k 1 N k U) µ(p N k 1 N k UP M l 1 M l ), with P j1 j 2 α = (0,..., 0, α j1 + 1,..., α j2, 0,..., 0).

20 Main theorem (Adcock, Hansen, P & Roman, 13) For ɛ > 0 and 1 k r, 1 N k N k 1 m k (log(ɛ 1 ) + 1) ( r ) µ N,M [k, l] s l log (N), and m k ˆm k (log(ɛ 1 ) + 1) log (N), where ˆm k satisfies r ( ) Nk N k µ N,M [k, l] s k, l = 1,..., r, ˆm k k=1 r k=1 s k s, s k max{ P N k 1 N k Uξ 2 : ξ = 1, ξ is (M, s)-sparse}. Suppose that ξ is a minimizer of min η 1 subject to P ΩN,m Uη P ΩN,mˆf δ. η C N l=1 Then, with probability exceeding 1 sɛ, we have that ξ x C (δ κ (1 + L s ) + σ M,s (f ) ), { Nk N k 1 log(ɛ 1 ) log(4κn s) for constant C, s := r k=1 s k, κ = max 1 k r m k }, L = 1 + and σ M,s (f ) is the best l 1 approximation of f by an (M, s)-sparse vector.

21 Fourier sampling and wavelet recovery (Adcock, Hansen, P & Roman, 13) If the sampling levels and sparsity levels correspond to wavelet scales, then the number of samples required at the k th Fourier sampling level is m k Log factors (s k + j k s j A j k ) where A is some constant dependent on the smoothness and the number of vanishing moments of the wavelet basis.

22 The test phantom Test phantom constructed by Guerquin-Kern, Lejeune, Pruessmann, Unser, 12

Thus, multi-level sampling only reaps their benefits for large values of N and

23 Resolution dependence (5% samples, varying resolution) Asymptotic sparsity and asymptotic incoherence are only witnessed when N is large. Thus, multi-level sampling only reaps their benefits for large values of N and the success of compressed sensing is resolution dependent. 256x256 Error: 19.86% 512x512 Error: 10.69%

24 Resolution dependence (5% samples, varying resolution) 1024x1024 Error: 7.35% 2048x2048 Error: 4.87% 4096x4096 Error: 3.06%

25 The optimal sampling strategy is signal dependent Existing theory on compressed sensing has been based constructing one sampling pattern for the recovery of all s-sparse signals. However, recall our conditions 1 N k N k 1 m k (log(ɛ 1 ) + 1) 1 r k=1 ( r ) µ N,M [k, l] s l log (N), l=1 ( ) Nk N k 1 1 µ N,M [k, l] s k, l = 1,..., r, ˆm k Clearly, our sampling pattern should depend both on the signal structure and local incoherences.

26 The optimal sampling strategy is signal dependent We aim to recover wavelet coefficients x from 10% Fourier samples. Original image (1024x1024) Signal structure of x Truncated (max = ) x10 5

27 The optimal sampling strategy is signal dependent For some multi-level sampling scheme, Ω N,m By solving Reconstruction min η 1 subject to η C N P ΩN,m Uη P ΩN,mˆf δ

28 The optimal sampling strategy is signal dependent We now flip the coefficient vector x, then x flip has the same sparsity of x but different signal structure. Signal structure of x flip If only sparsity matters Truncated (max = )...then using the same sampling pattern Ω N,m min η 1 subject to η C N P ΩN,m Uη P ΩN,mˆf flip δ will recover x flip. So we can recover x = (x flip ) flip x10 5

29 The optimal sampling strategy is signal dependent The reconstruction So signal structure matters... This observation of signal dependence opens up the possibility of designing optimal sampling patterns for specific types of signals.

30 Outline Introduction A new theory for compressed sensing (Joint work with Ben Adcock, Anders Hansen and Bogdan Roman) Compressed sensing with non-orthonormal systems

31 A more general problem We have explained the use of multi-level sampling schemes for orthonormal systems. But... Our sparsifying dictionary could be a frame (e.g. shearlets), or we may want to minimize with a TV norm. Our sampling vectors need not form an orthonormal basis (e.g. nonuniform Fourier samples).

32 A more general problem We aim to extend the theory to understand how local incoherence and local signal structures affect the minimizers of min Dη 1 subject to P ΩN,m Uη P ΩN,mˆf δ η C N where ˆf = Ux, U C M N, D C d N. We will focus on the following two examples: Case I: D is the finite differences matrix and U is the Fourier matrix. Case II: D = I and U is the matrix for shearlets with Fourier sampling.

33 Co-sparsity (Introduced by Elad et al, 11) min η C N Dη 1 subject to P ΩN,m Uη P ΩN,mˆf δ If the rows of D are highly redundant, then Dx cannot be sparse unless x = 0 = focus on the zeros of Dx. The co-sparse set is the index set Λ for which P Λ Dx = 0. Let dim Null(P Λ D) = s. Let {w j : j = 1,..., s} be an orthonormal basis for Null(P Λ D) and W Λ = (w 1 w 2... w s ),

34 Assumptions and incoherence min η C N Dη 1 subject to P ΩN,m Uη P ΩN,mˆf δ. We require that X := W Λ (W Λ U UW Λ ) 1 W Λ exists. (C inv ) The following conditions determine the types of signals that can be recovered (cf. notions of identifiability by Fuchs 04, Tropp 04, Peyré 11): (D P Λ ) U UXW Λ < 1 (C id 1) inf u Null(D P Λ ) (D P Λ ) D P Λ c sgn(p Λ c Dx) u < 1. (C id 2) Coherence is between U and Null(P Λ D) = span {w j : j = 1,..., s}: let µ N,M [k] = max { µ N,M (UW Λ )[k], µ N,M (UXW Λ )[k], µ N,M (U(P Λ D) )[k] } µ N,M [k, j] = µ N,M [k] max {µ N,M (UW Λ )[k, j], µ N,M (UXW Λ )[k, j]}

35 for constant C, s := r k=1 s N k and κ = max k N k 1 1 k r m k. Recovery statement (P. 13) Let ɛ > 0 and ˆf = Ux. Recall X := W Λ (W Λ U UW Λ ) 1 W Λ. For m k log(ɛ 1 + 1) log(kn s) (N k N k 1 ) and m k log(ɛ 1 + 1) log(kn s) ˆm k with 1 r k=1 r k=1 s k UX 2 s, Suppose that ξ is a minimizer of min η C N Dη 1 r µ N,M [k, j] s j, j=1 ( ) Nk N k 1 1 µ N,M [k] s k ˆm k s k max{ P N k 1 N k UXW Λ ξ 2 : ξ = 1}. subject to P ΩN,m Uη P ΩN,mˆf δ. Then, with probability exceeding (1 sɛ), ξ x C (PΛ D) (( ) 1 2 X + s L κ ) δ + P Λ Dx 1.

36 Case I: Total variation with Fourier samples U is the unitary Discrete Fourier matrix, D = Given Λ, if Λ c = {γ j : j = 1,..., s 1} with γ 0 = 0, γ s = 2 n, then Null(P Λ D) = span { (γ j γ j 1 ) 1/2 χ (γj 1,γ j ] : j = 1,..., s }. (C inv ) holds, (C id 1) is trivial and (C id 2) holds if there is no stair-casing: j s.t. (P Λ c Dx) j = (P Λ c Dx) j+1 = ±1. (Peyré et al, 2011) ( ( { }) 1 1 Lj 1 µ N,M [k] = O N k 1 ), µ N,M [k, j] = O min N k 1, N k 1 2 where n L j is the shortest length of the support of vectors in level j.

37 Case II: Shearlet reconstructions from Fourier samples min η C N η 1 subject to P ΩN,m U df Vdsη P ΩN,mˆf δ. where V ds is some (tight) discrete shearlet transform. Let a j be the j th row of V ds. Λ c =: indexes the sparse shearlet representation. We can assume that {a j : j } is a linearly independent set. In particular, X = P (P U UP ) 1 P = P (P V ds V ds P ) 1 P exists and (C inv ) holds. (C id 2) is trivial and (C id 1) is true whenever sup i j a i, a j + max i j,j i a i, a j < 1. Note that the Gram matrices of shearlets/curvelets are known to have strong off diagonal decay properties (Grohs & Kutyniok, 2012 ). µ(p K U) = O ( K 1), µ(up K ) = O ( K 1).

38 Example min kηk1 subject to η CN PΩN,m Udf V 1 η PΩN,m fˆ δ. 6.25% subsampling map (2048x2048) Courtesy of Anders Hansen and Bogdan Roman Image (2048x2048)

39 Example DB-4 reconstruction (2048x2048) Shearlet reconstruction (2048x2048) Courtesy of Anders Hansen and Bogdan Roman

40 Example Curvelet reconstruction (2048x2048) Contourlet reconstruction (2048x2048) Courtesy of Anders Hansen and Bogdan Roman

41 Conclusions There is a gap between the theory and the use of compressed sensing in many real world problems. By introducing notions of asymptotic incoherence, asymptotic sparsity and multi-level sampling, we can explain the success of variable density sampling schemes. Two key consequences of our theory: (1) Compressed sensing is resolution dependent. (2) Successful recovery is signal dependent, thus, an understanding of local incoherence and sparsity patterns of certain types of signals can lead to optimal sampling patterns. These ideas are applicable to non-orthonormal systems, including frames and total variation. Breaking the coherence barrier: asymptotic incoherence and asymptotic sparsity in compressed sensing. Adcock, Hansen, Poon & Roman 13 Not covered in this talk: Extension to infinite dimensional framework We recovered x C N from Ux, U C N N. But the MRI problem samples the continuous Fourier transform and is infinite dimensional. Direct application of finite dimensional methods results in artefacts.

On the coherence barrier and analogue problems in compressed sensing

On the coherence barrier and analogue problems in compressed sensing Clarice Poon University of Cambridge June 1, 2017 Joint work with: Ben Adcock Anders Hansen Bogdan Roman (Simon Fraser) (Cambridge)