ELEG Compressive Sensing and Sparse Signal Representations

Transcription

1 ELEG Compressive Sensing and Sparse Signal Representations Gonzalo R. Arce Depart. of Electrical and Computer Engineering University of Delaware Fall 211 Compressive Sensing G. Arce Fall, / 42

2 Outline Compressive sensing signal reconstruction Geometry of CS Reconstruction Algorithms Matching Pursuit Orthogonal Matching Pursuit Iterative Hard Thresholding Weighted Median Regression Compressive Sensing G. Arce Fall, / 42

3 Compressive Sensing Signal Reconstruction Goal: Recover signal x from measurements y Problem: Random projectionφnot full rank (ill-posed inverse problem) Solution: Exploit the sparse/compressible geometry of acquired signal x y Φ x Compressive Sensing G. Arce CS Signal Reconstruction Fall, / 42

4 Suppose there is a S-sparse solution to y = Φx Combinatorial optimization problem Convex optimization problem (P) min x x s.t. Φx = y (P1) min x x 1 s.t. Φx = y If 2δ 3s +δ 2s 1, the solutions(p) and (P1) are the same. E. Candès. Compressive Sampling. Proc. Intern. Congress of Math., China, April 26. Compressive Sensing G. Arce CS Signal Reconstruction Fall, / 42

5 The Geometry of CS Sparse signals have few non-zero coefficients X 1 X 1 R 3 R 3 X 2 X 2 X 3 1-sparse X 3 2-sparse Compressive Sensing G. Arce Geometry of CS Fall, / 42

6 l Recovery Reconstruction should be Consistent with the model: x as sparse as possible min x Consistent with the measurements: y = Φx X 1 R 3 x X 2 X 3 x:y= Φx Compressive Sensing G. Arce Geometry of CS Fall, / 42

7 l Recovery min x x s.t. Φx = y x number of nonzero elements Sparsest signal consistent with the measurements y Requires only M << N measurements Combinatorial NP-hard problem: for x R N with sparsity S, the complexity is O(N S ) Too slow for implementation Compressive Sensing G. Arce Geometry of CS Fall, / 42

8 l 1 Recovery A more realistic model min x x 1 s.t. Φx = y Thel 1 norm also induces sparsity. The constraint is given by y = Φx. X 1 R 3 x X 2 X 3 x:y= Φx Compressive Sensing G. Arce Geometry of CS Fall, / 42

9 Phase Transition Diagram In thel 1 minimization, there is a defined region on (M/N, S/M) which ensures successful recovery δ = M/N is a normalized measure of the problem indeterminacy. ρ = S/M is a normalized measure of the sparsity. Red region - unsuccessful recovery or exact reconstruction typically fails. Blue region - successful recovery or exact reconstruction typically occurs. Compressive Sensing G. Arce Geometry of CS Fall, / 42

10 l 2 Recovery min x x 2 s.t. Φx = y Least square solution x = (Φ T Φ) 1 Φ T y Solved by using quadratic programming: Least squares solution Interior-point methods Solution is almost never sparse X 1 R 3 x' x X 2 X 3 x:y= Φx Compressive Sensing G. Arce Geometry of CS Fall, / 42

11 l 2 Recovery Problem: smalll 2 does not imply sparsity x x' x has smalll 2 norm but it is not sparse Compressive Sensing G. Arce Geometry of CS Fall, / 42

12 CS Example in the Time Domain y = Φx Random Projection Time domain sparse signal Nonlinear Reconstruction min x 1 s.t. y = Φx Error signal Compressive Sensing G. Arce Geometry of CS Fall, / 42

13 CS Example in the Wavelet Domain Reconstruction of an image (N = 1 megapixel) sparse (S = 25, ) in the wavelet domain from M = 96, incoherent measurements. Original (25K wavelets) Recovered Image E. J. Candès and J. Romberg Sparsity and Incoherence in Compressive Sampling. Inverse Problems. vol.23, pp Compressive Sensing G. Arce Geometry of CS Fall, / 42

14 Reconstruction Algorithms Different formulations and implementations have been proposed to find the sparsest x subject to y = Φx Difficult to compare results obtained by different methods Those are broadly classified in: Regularization formulations (Replace combinatorial problem with convex optimization) Greedy algorithms (Iterative refinement of a sparse solution) Bayesian framework (Assume prior distribution of sparse coefficients) Compressive Sensing G. Arce Reconstruction Algorithms Fall, / 42

15 Greedy Algorithms Iterative algorithms that select an optimal subset of the desired signal x at each iteration. Some examples: Matching Pursuit (MP) min x x s.t. Φx = y Orthogonal Matching Pursuit(OMP) min x x s.t. Φx = y Compressive Sampling Matching Pursuit (CoSaMP) min x y Φx 2 2 s.t. x S Iterative Hard Thresholding (ITH) min x y Φx 2 2 s.t. x S Weighted Median Regression min x y Φx 1 +λ x Those methods build an approximation of the sparse signal at each iteration. Compressive Sensing G. Arce Reconstruction Algorithms Fall, / 42

16 Matching Pursuit (MP) Sparse approximation algorithm used to find the solution to Key ideas: min x x s.t. Φx = y Measurements y = Φx are composed of sum of S columns ofφ. The algorithm needs to identify which S columns contribute to y. y x S Compressive Sensing G. Arce Matching Pursuit Fall, / 42

17 Matching Pursuit (MP) Algorithm Input and Observation vector y Initialize Measurement MatrixΦ Parameters Initial Solution x () = Initial residual r () = y Iteration k Compute the current error: c (k) j = φ j, r (k) Identify the index ĵ such that: ĵ = max j c (k) Update x (k) ĵ = x (k 1) ĵ + c (k) ĵ Update r (k) = r (k 1) c (k) φĵ ĵ j D. Donoho et. al, SparseLab Version Compressive Sensing G. Arce Matching Pursuit Fall, / 42

18 MP Reconstruction Example reconstruc. signal Iteration residue - Iteration reconstruc. signal Iteration residue - Iteration 5 Compressive Sensing G. Arce Matching Pursuit Fall, / 42

19 MP Reconstruction Example Original Signal reconstruc. signal Iteration 32 8 x Residue Iteration 32 Compressive Sensing G. Arce Matching Pursuit Fall, / 42

20 Orthogonal Matching Pursuit (MP) Sparse approximation algorithm used to find the solution to Key ideas: min x x s.t. Φx = y Measurements y = Φx are composed of sum of S columns ofφ. The algorithm needs to identify which S columns contribute to y. y x S Compressive Sensing G. Arce Orthogonal Matching Pursuit Fall, / 42

21 Orthogonal Matching Pursuit (OMP) Algorithm Input and Observation vector y Initialize Measurement Matrix Φ Parameters Initial Solution x () = Initial Solution Support S () =support{ x () } = Initial residual r () = y Iteration k Compute the current error: c (k) j = φ j, r (k) Identify the index ĵ such that: ĵ = max j c (k) j Update the support S (k) =S (k 1) ĵ Update the matrixφ S (k) = [,...,φ i,...,,...,φ j,..., ] Update the solution x (k) = (Φ T Φ S (k) S (k)) 1 Φ T y S (k) Update r (k) = y Φ S (k) x (k) D. Donoho et. al, SparseLab Version Compressive Sensing G. Arce Orthogonal Matching Pursuit Fall, / 42

22 Remarks on the OMP Algorithm S (k) is the support of x at the iteration k and Φ S (k) is the matrix that contains the columns fromφthat belongs to this support. The updated solution gives the x (k) that solves the minimization problem Φ S (k)x y 2 2. The solution is given by Φ T S (k) (Φ S (k)x y) = Φ T S (k) r (k) = This solution shows that the columns inφ T S (k) are orthogonal to r (k). Compressive Sensing G. Arce Orthogonal Matching Pursuit Fall, / 42

23 OMP Reconstruction Example reconstruc. signal Iteration residue - Iteration reconstruc. signal Iteration residue - Iteration 5 Compressive Sensing G. Arce Orthogonal Matching Pursuit Fall, / 42

24 OMP Reconstruction Example reconstruc. signal Iteration residue - Iteration x reconstruc. signal Iteration residue - Iteration 1 Compressive Sensing G. Arce Orthogonal Matching Pursuit Fall, / 42

25 Iterative Hard Thresholding Sparse approximation algorithm that solves the problem min x y Φx 2 2 +λ x. Iterative algorithm that computes a descent direction given by the gradient of thel 2 -norm. The solution at each iteration is given by finding: x (k+1) = H λ (x (k) +Φ T (y Φx (k) )) H λ is a non-linear operator that only retains the coefficients with the largest magnitude. T. Blumensath and M. Davies. Iterative Hard Thresholding for Compressed Sensing. Journal of Appl. Comp. Harm. Anal. vol. 27, no. 3, pp , 29. Compressive Sensing G. Arce Iterative Hard Thresholding Fall, / 42

26 Iterative Hard Thresholding Input and Observation vector y Initialize Measurement Matrix Φ Parameters x () = Iteration k Compute an update: a (k+1) = x (k) +Φ T (y Φx (k) )) Select the largest elements: x (k+1) = H λ (a (k+1) ). Pull other elements toward zero. D. Donoho et. al, SparseLab Version Compressive Sensing G. Arce Iterative Hard Thresholding Fall, / 42

27 ITH Reconstruction Example reconstruc. signal Iteration residue - Iteration reconstruc. signal Iteration residue - Iteration 5 Compressive Sensing G. Arce Iterative Hard Thresholding Fall, / 42

28 ITH Reconstruction Example Original Signal reconstruc. signal Iteration 1 8 x residue - Iteration 1 Compressive Sensing G. Arce Iterative Hard Thresholding Fall, / 42

29 Weighted Median Regression Reconstruction Algorithm When the random measurements are corrupted by noise the most widely used CS reconstruction algorithms solve the problem min x y Φx 2 2 +τ x 1. whereτ is the regularization parameter that balances the conflicting task of minimizing thel 2 norm while yielding a sparse solution. Asτ increases the solution is sparser. Asτ decreases the solution approximates to least square solution. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

30 In the solution to the problem min x y Φx 2 2 +τ x 1. Thel 2 term is the data fitting term, induced by the Gaussian assumption of the noise contamination. Thel 1 term is the sparsity promoting term. However, thel 2 term leads to the least square solution which is known to be very sensitive to high-noise level and outliers present in the contamination. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

31 Example of signal reconstruction when outliers are present in the random projections Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

32 To mitigate the effect of the impulsive noise in the compressive measurements, a more robust norm for the data-fitting term is min y Φx 1 x }{{} +λ x }{{} Data fitting Sparsity (1) LAD offers robustness to a broad class of noise, in particular to heavy tail noise. Optimum under the maximum likelihood principle when the underlying contamination follows a Laplacian-distributed model. J. L. Paredes and G. Arce. Compressive Sensing Signal Reconstruction by Weighted Median Regression Estimate. ICASSP 21. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

33 Drawbacks: Solvingl -LAD optimization is N p -hard problem. Direct solution is unfeasible even for modest-sized signal. To solve this multivariable minimization problem: Solve a sequence of scalar minimization subproblems. Each subproblem improves the estimate of the solution by minimizing along a selected coordinate with all other coordinates fixed. Closed form solution exists for the scalar minimization problem. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

34 Signal reconstruction x = min x M (y Φx) i +λ x i=1 The solution using Coordinate Descent Method is given by x n = min x n M N N y i φ i,j x j φ i,n x n +λ x n + λ x j i=1 j=1,j n j=1,j n x n = min x n M φ in y i N j=1,j n φ i,jx j x n +λ x n φ }{{ in }{{} } i=1 Weighted least absolute deviation Q(x n) Regulariz. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

35 The solution to the minimization problem is given by x n = MEDIAN( φ in y i N At k th iteration: x (k) n = x n = min x n Q(x n )+λ x n j=1,j n φ i,jx j φ in { xn, if Q( x n )+λ Q(), otherwise ) M i=1; n = 1, 2,..., N Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

36 The iterative algorithm detects the non-zero entries of the sparse vector by Robust signal estimation: computing a rough estimate of the signal that minimizes the LAD by the weighted median operator. Basis selection problem: applying a hard-threshold operator on the estimated values. Removing the entry s influence from the observation vector. Repeating these steps again until a performance criterion is reached. Original Signal Iteration Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

37 Input and Observation vector y Initialize Measurement Matrix Φ Parameters Number of Iterations K Iteration counter: k = 1 Hard-thresholding value: λ = λ i Estimation at k = 1, x () = Iteration Step A For the n-th entry { of x, n = 1, 2,..., N compute x n = MEDIAN φ in yi N j=1;j n φijˆxj Step B Step C Output x (k) n = } φ in k i=1 { xn, if Q( x n )+λ Q(), otherwise, Update the hard-thresholding parameter and the estimation of x x (k+1) = x (k) λ = λ i β k Check stopping criterium If k K then set k = k+1 and go to Step A; otherwise, end. Recovered sparse signal ˆx Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

38 Example Signal Reconstruction from Projections in Non-Gaussian Noise CoSaMP L1-Ls Original signal and noisy projections Matching Pursuit Weighted Median Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

39 Performance Comparisons Exact reconstruction for noiseless signals (normalized error is smaller than.1% of the signal energy). Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

40 Performance Comparisons SNR vs. Sparsity for different types of noise Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

41 Convex Relaxation Approach: Thel -LAD problem can be relaxed by solving the following problem min y Φx 1 x }{{} +λ x 1 }{{}. (2) Data fitting Sparsity This problem is convex and can be solved by linear programming. l 1 -norm may not induce the necessary sparsity in the solution. It requires more measurements to achieve a target reconstruction error. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42

42 The solution for the problem x = min x M y i Φ i x +λ x 1 i=1 is given by ˆx = MEDIAN(λ, φ 1 y 1 φ 1,..., φ M y M φ M ). The regularization process induced by thel 1 -term leads to a weighting operation with the zero-valued sample in the WM operation. Large values of the regularization parameter implies large value for the weight corresponding to the zero-valued sample (this favors sparsity). The solution is biased. Compressive Sensing G. Arce Weighted Median Regression Fall, / 42