Reconstruction from Fourier measurements using compactly supported shearlets Jackie Ma Technische Universität Berlin Department of Mathematics Straße des 17. Juni 136, 10623 Berlin Email: ma@math.tu-berlin.de Abstract—We study the reconstruction problem of a compactly supported function from its Fourier coefficients using a compactly supported shearlet system. We assume that only finitely many Fourier samples are accessible and the reconstruction can only be constructed using finitely many shearlets. Our main result shows that stable recovery of the signal is possible provided the number of measurements is up to a constant equal to the number of shearlet elements that are used for the approximation. Numerical experiments with MR data are presented which emphasize the advantages of shearlets compared to other systems such as wavelets. I. I NTRODUCTION A general task in sampling theory is the reconstruction of an object from finitely many measurements. Such a problem also appears in signal processing and medical imaging the latter being a main motivation of this paper with view to applications. Further, we model the reconstruction problem in a Hilbert space H with inner product h·, ·i. The measurements are assumed to be acquired with respect to a fixed sampling system {s1 , s2 , . . .} ⊂ H and the reconstruction is computed with respect to another system {r1 , r2 , . . .} ⊂ H which is assumed to form a frame. Recall that a set of vectors (ϕi )i∈I ⊂ H is called a frame for H if there exist 0 < A ≤ B < ∞ such that X Akf k2 ≤ |hf, ϕi|2 ≤ Bkf k2 , ∀ f ∈ H. i∈I The concrete reconstruction problem is then formulated as follows: Given finitely many linear measurements hf, s1 i, . . . , hf, sM i, M ∈ N of a function f ∈ H, find a reconstruction fN ∈ span{r1 , . . . , rN }, N ∈ N that is numerically stable and whose error converges to zero as N tends to infinity. Reconstruction problems of this type have already been studied, e.g. by Unser and Aldroubi in [1] and later also by Eldar in [2], [3] for the case N = M which is known as consistent reconstruction. We study this reconstruction problem by using generalized sampling (GS) which can be seen as an extension of consistent reconstruction. II. G ENERALIZED SAMPLING Generalized sampling was introduced by Adcock et al. in [4], [5], [6] and studies the reconstruction problem explained c 978-1-4673-7353-1/15/$31.00 2015 IEEE above given that the sampling system and reconstruction system form frames. The key difference between generalized sampling and consistent reconstruction is to allow the number of measurements M to vary independently of the number of reconstruction elements N . This flexibility enables generalized sampling to overcome some of the barriers of consistent reconstruction [4]. A. GS reconstruction method Given a sampling system {s1 , s2 , . . .} ⊂ H we define the sampling space S = span{sk : k ∈ N} ⊂ H and the finite dimensional version of this sampling space is denoted by SM = span{s1 , . . . , sM }, M ∈ N. Analogously for a reconstruction system {r1 , r2 , . . .} ⊂ H we define the reconstruction space R = span{rk : k ∈ N} ⊂ H and, likewise, its finite dimensional version is defined as RN = span{r1 , . . . , rN }, N ∈ N. We will assume the sampling system to form an orthonormal basis and the reconstruction system to be a frame, since this will exactly be our setup, see Section IV. Further it is natural to assume a certain subspace condition, indeed, we require R ∩ S ⊥ = {0} and R + S is closed. (1) This guarantees a well posedness of the finite dimensional reconstruction problem, cf. Theorem II.1. Theorem II.1 ([6]). For any N, M ∈ N let SM and RN be as above and PSM be the following finite rank operator PSM : H → SM , f 7→ M X hf, sk isk . k=1 Let N ∈ N be fixed. If (1) holds, then there exists an M ∈ N such that the system of equations hPSM fN,M , rj i = hPSM f, rj i, j = 1, . . . , N (2) has a unique solution fN,M ∈ RN . Moreover, the smallest M ∈ N such that the system is uniquely solvable is the least number M0 ∈ N so that cN,M := inf kPSM f k > 0. f ∈RN , kf k=1 (3) Furthermore, kf − PRN f k ≤ kf − fN,M k ≤ A. Assumptions on the generator 1 cN,M kf − PRN f k, (4) where PRN : H → RN denotes the orthogonal projection onto RN . Let φ and ψ be compactly supported functions in L2 (R2 ) e 1 , x2 ) := ψ(x2 , x1 ). We assume there exist and define ψ(x some constants C1 , C2 > 0 such that 1 , (1 + |ξ1 |)r (1 + |ξ2 |)r min{1, |ξ1 |α } b 1 , ξ2 )| ≤ C2 · |ψ(ξ . (1 + |ξ1 |)r (1 + |ξ2 |)r b 1 , ξ2 )| ≤ C1 · |φ(ξ Definition II.2 ([6]). The solution fN,M in Theorem II.1 is called generalized sampling reconstruction. Definition II.3 ([1], [7]). The quantity cN,M in (3) is called the infimum cosine angle between the subspaces RN and SM . B. Stable sampling rate In light of Theorem II.1 it is crucial to have control over cN,M . This motivates the definition of the stable sampling rate. Definition II.4. For any fixed N ∈ N and θ > 1 the stable sampling rate Θ(N, θ) is defined as 1 Θ(N, θ) = min M ∈ N : cN,M > . θ The stable sampling rate determines the number of measurements M that are needed in order to find an approximation fN ∈ RN or equivalently, find N coefficients, such that the angle cN,M is controlled by the threshold θ. III. S HEARLETS Shearlet systems were first introduced by K. Guo, G. Kutyniok, D. Labate, W.-Q Lim and G. Weiss in [8], [9] and the following notations have became standard in shearlet theory. The parabolic scaling matrices with respect to scale j ∈ N ∪ {0} are denoted by j j/2 2 0 2 0 e A2j := , A2j := , 0 2j/2 0 2j and the shearing matrices with parameter k ∈ Z are 1 k Sk = . 0 1 These operations, together with the standard integer shift of functions in L2 (R2 ) are then used to define the cone adapted discrete shearlet system. Definition III.1 ([10]). Let φ, ψ, ψe ∈ L2 (R2 ) be the generating functions and c = (c1 , c2 ) ∈ (R+ )2 . Then the (cone adapted discrete) shearlet system is defined as e c) SH(φ, ψ, ψ, = {φ(· − c1 m) : m ∈ Z2 } n o ∪ 23j/4 ψ ((Sk A2j ) · −cm) : j ≥ 0, |k| ≤ 2j/2 , m ∈ Z2 n o e2j ) · −e ∪ 23j/4 ψe (SkT A cm : j ≥ 0, |k| ≤ 2j/2 , m ∈ Z2 where the multiplication of c and e c with the translation parameter m should be understood pointwise. Under some regularity condition, see (5) and (6) in Subsection III-A, the shearlet system defined above can form a frame and an explicit construction of such systems and their generators φ, ψ is given in [10]. (5) (6) where the regularity parameters α, r > 0 are large enough so that the shearlet system gives rise to a frame for L2 (R2 ), [10]. B. Assumptions on finite dual shearlets In [11] it was shown that the frame operator for continuous shearlets acts as a Fourier multiplier and the existence of a dual shearlet that has the same asymptotic Fourier decay as the original shearlet was given. In fact, similar computations as those in Lemma 4.2 in [11] can be performed for discrete shearlets as defined in Definition III.1 for appropriate sampling constants c. In the following we make this result more precise for our setup. Let (ψλ )λ∈N be a compactly supported shearlet frame for H = L2 (R2 ) and for N ∈ N let HN := span{ψλ : λ ≤ N }. Note that (ψλ )λ forms a frame for HN . Now we assume that there exist finite dual shearlets (ψλd,N )λ≤N that asymptotically have the same behavior as the original shearlet atoms (ψλ )λ≤N , i.e. we assume cλ (ξ)| as N → ∞ |(ψλd,N )∧ (ξ)| |ψ (7) for ξ ∈ R2 . Note that by definition of a dual frame, the finite dual shearlet system (ψλd,N )λ≤N can be used to represent any f ∈ HN by X X f= hf, ψλd,N iψλ = hf, ψλ iψλd,N , λ≤N λ≤N see [12] for the definition of dual frames. Assumption (7) is, for instance, fulfilled for digital shearlets [13]. In fact, for these shearlets, dual shearlets are given by d b d φ(ξ) ψ ψej,k (ξ) j,k (ξ) [ [ d,N d,N [ d,N (ξ) = , ψj,k (ξ) = , ψej,k (ξ) = , φ Γ(ξ, J) Γ(ξ, J) Γ(ξ, J) where X X d 2 2 d e |ψj,k (ξ)| + |ψj,k (ξ)| b 2+ Γ(ξ, J) = |φ(ξ)| j≤J k≤2j/2 is bounded from above and below, see [14] and the references therein. NJ indicates that only shearlets up to a certain scale J are considered, cf. Subsection IV-A. Further, it was shown in [10] that X X X X d 2 b + d A ≤ φ(ξ) |ψ |ψej,k (ξ)|2 ≤ B j,k (ξ)| + j≥0 k≤2j/2 j≥0 k≤2j/2 where A and B are the lower and upper frame bounds of the full shearlets system (ψλ )λ∈N , respectively. Therefore the assumption made in (7) can be justified. We wish to mention that recently a new type of shearlet systems has been introduced in [15] where a corresponding dual shearlet can be given explicitly. These dual shearlets corresponding to the new dualizable shearlets also show the same asymptotic decay as the dualizable shearlet generators similar to (7). Thus, emphasizing the rationality of such assumptions. IV. S TABLE SHEARLET RECONSTRUCTIONS FROM F OURIER MEASUREMENTS Finally, for M = (M1 , M2 ) ∈ N × N let o n (ε) (ε) SM = span s` : ` = (`1 , `2 ) ∈ Z2 , |`i | ≤ Mi , i = 1, 2 be the finite dimensional sampling space. Note that M = (M1 , M2 ) determines the size of the grid and thus, the total number of possible measurements are asymptotically of order M1 · M2 . In the event that the stable sampling rate grows linearly we would have M1 · M2 = O(22J ) as for the wavelet case shown in [16], [17]. A. Shearlet reconstruction and Fourier sampling space B. Almost linear stable sampling rate Without loss of generality we can assume that the generating scaling function φ and the shearlets ψ, ψe are compactly supported in [0, a]2 where a is some positive integer. Then we consider all scaling functions whose support intersect [0, a]2 and denote this index set by Ω, i.e. Ω = m ∈ Z2 : supp φm ∩ [0, a]2 6= ∅ . Theorem IV.1 shows that the angle between the shearlet reconstruction space and the Fourier sampling space can be controlled with an almost linear sampling ratio. Similarly, we consider all shearlets whose support intersect [0, a]2 . For this, let J − 1 ∈ N ∪ {0} be a fixed scale. Then by ΛJ we denote the following paramater set n o ΛJ = (j, k, m) : 0 ≤ j ≤ J − 1, |k| ≤ 2j/2 , m ∈ Ωj,k , where Ωj,k = {m ∈ Z2 : supp ψj,k,m ∩ [0, a]2 6= ∅}. Note that Ωj,k is of finite cardinality. Analogously we set n o e J = (e ee e Λ j, e k, m) e : 0≤e j ≤ J − 1, |e k| ≤ 2j/2 , m e ∈Ω j,k 2 e e e = {m with Ω e ∈ Z2 : supp ψeej,ek,m j,k e ∩[0, a] 6= ∅} and define RNJ := span {φm : m ∈ Ω} ∪ {ψj,k,m : (j, k, m) ∈ ΛJ } n o e eJ e ∪ ψeej,ek,m : ( j, k, m) e ∈ Λ . (8) e e be a compactly supTheorem IV.1 ([18]). Let SH(φ, ψ, ψ) e We further ported shearlet frame with generators φ, ψ, and ψ. assume the generators to be as in Subsection III-A and with (finite) duals as in Subsection III-B. Let N ≤ NJ = O(22J ). Then for all θ > 1 there exists Sθ > 0 such that 1 ε fk ≥ , as N → ∞ cN,M = inf kPSM f ∈RN θ kf k=1 where M = (M1 , M2 ) ∈ N × N obeys Mi = dSθ · 2J(1+δ) /εe, i = 1, 2 2 with δ ≥ 2r−1 and r > 0 is the regularity parameter from (5) and (6). Further, the constant Sθ does not depend on N but on θ, ω, α, and r. Remark IV.2. The dependence of the constant Sθ on all other relevant variables can be explicitly computed. Moreover, without the assumptions made in Subsection III-A, the result holds true with Sθ depending on the lower frame bound AN of the finite shearlet frame (ψλ )λ≤N . Up to a fixed scale J − 1 we, asymptotically, have NJ = V. N UMERICAL EXPERIMENTS O(22J ) many generating functions in RNJ as J → ∞. Note In this section we provide some reconstructions from MR that by construction at each scale we only have finitely many data using complex exponentials (Fourier inversion), wavelets elements, therefore, an ordering can be performed quite natuand compactly supported shearlets. Although our main result rally namely we order the system along scales and within the guarantees stable and convergent reconstructions when the 2 scales, the translations in Z are ordered in a lexicographical sampling rate is almost linear it is not efficient (and also not manner. By R we denote the reconstruction space that contains all necessary) to acquire that many samples in practice. Thus, we will subsample the Fourier data according to some common shearlets across all scales, i.e. subsampling patterns used in applications such as spirals and R = span {φm : m ∈ Ω} ∪ {ψj,k,m : (j, k, m) ∈ ΛJ , J ≥ 0} radial lines, see Figure 2a and Figure 2b. A. Data n o e eJ , J ≥ 0 e ∪ ψeej,ek,m : ( j, k, m) e ∈ Λ . The underlying Fourier data, also called k-space, was e acquired by a multi-channel acquisition consisting of four To define the sampling space we first choose T1 , T2 > 0 so channels. All four k-spaces had a fixed resolution of 128×128 1 that R ⊂ L2 ([−T1 , T2 ]2 ). Let ε ≤ T1 +T < 1 determine the which results in an 128 × 128 image in time by applying an 2 sampling density. Then we define the sampling vectors on a inverse Fourier transform, see Figure 1a. (ε) uniform grid by s` = εe2πiεh`,·i , ` ∈ Z2 over [−T1 , T2 ]2 and The single channel images are combined by using the sumthe sampling space S by of-squares method, [19]. The resulting sum-of-squares image n o from all four channels of the original k-space is used as the (ε) S (ε) = span s` : ` ∈ Z2 . reference image and is depicted in Figure 1. 20 20 20 40 40 40 60 60 60 80 80 80 100 100 100 120 120 120 50 100 150 200 250 (a) 4 channel reconstructions (b) Combined image by sumof-squares Fig. 1: Reference images computed with complete data B. Methods We mimic an acceleration factor of almost 5 by using a 0 − 1 mask to subsample the k-space, cf. Fig. 2a and Fig. 2b. We then compute each channel image separately and combine them by sum-of-squares. The following problems and methods are used to obtain the single channel images • Shearlets: the reconstructions are obtained by solving the following standard analysis formulation β e − bk22 + λkSH(u)k1 , min kFu u 2 where Fe is the subsampled Fourier operator, b is the subsampled input data, SH is the shearlet transform, β and λ are some parameters larger than zero, and u is the image. For our numerical tests we used the shearlab package available at (a) Phyllotaxis spiral mask: 20.37 % (b) Radial mask: 20.74% 20 20 40 40 60 60 80 80 100 100 120 120 (c) Reconstruction: shearlets, spiral (d) Reconstruction: shearlets, radial 20 20 40 40 60 60 80 80 100 100 120 120 (e) Reconstruction: wavelets, spiral (f) Reconstruction: wavelets, radial 20 20 40 40 60 60 http://www.shearlab.org • A three scale shearlet transform was used with number of directions set to [0 0 1], see http://www.shearlab.org and also [23], [20] for more interest in the algorithmic realization of shearlets. Wavelets: reconstructions are computed using the SparseMRI package from Lustig et al. [21] which is downloaded from http://www.eecs.berkeley.edu/˜mlustig/Software.html • Fourier: computed by simple Fourier inversion of the subsampled data. C. Results The shearlet reconstructions sare shown in Figure 2 and the relative errors are given in Table I. As we can see, the shearlet reconstruction show reduced artifacts without loosing fine details. This can be explained in terms of the sparse approximation rate of cartoon-like-functions [22], more precisely, shearlets provide an almost optimal rate of order N −2 for cartoon-like functions, while wavelets only reach N −1 and Fourier N −1/2 , respectively, see [23]. Relative error kf − fN k/kf k Shearlets Wavelets Phyllotaxis Spiral 0.0653 0.1145 Radial 0.0914 0.1531 Fourier 0.2570 0.2733 TABLE I: Relative error for Figure 2 80 80 100 100 120 120 (g) Reconstruction: Fourier inversion, spiral (h) Reconstruction: Fourier inversion, radial Fig. 2: Reconstructions using a spiral and radial mask. ACKNOWLEDGMENT The author would like to thank Gitta Kutyniok for helpful discussions. Moreover, the author thanks Dr. Sina Straub from German Cancer Research Center (DKFZ) for providing the MRI data and Dr. Matthias Dieringer from Siemens AG, Healthcare Sector for providing a code to read the raw data into MATLAB. Further, the author acknowledges support from the Berlin Mathematical School as well as the DFG Collaborative Research Center TRR 109 ”Discretization in Geometry and Dynamics”. R EFERENCES [1] M. Unser and A. Aldroubi, “A general sampling theory for nonideal acquisition devices.” IEEE Transactions on Signal Processing, vol. 42, no. 11, pp. 2915–2925, 1994. [2] Y. C. Eldar, “Sampling with arbitrary sampling and reconstruction spaces and oblique dual frame vectors,” J. Fourier Anal. Appl., vol. 9, no. 1, pp. 77–96, 2003. [3] Y. C. Eldar and T. Werther, “General framework for consistent sampling in Hilbert spaces,” Int. J. Wavelets Multiresolut. Inf. Process., vol. 3, no. 4, pp. 497–509, 2005. [4] B. Adcock and A. C. Hansen, “A generalized sampling theorem for stable reconstructions in arbitrary bases,” J. Fourier Anal. Appl., vol. 18, no. 4, pp. 685–716, 2012. [5] ——, “Stable reconstructions in Hilbert spaces and the resolution of the Gibbs phenomenon,” Appl. Comput. Harmon. Anal., vol. 32, no. 3, pp. 357–388, 2012. [6] B. Adcock, A. C. Hansen, and C. Poon, “Beyond consistent reconstructions: optimality and sharp bounds for generalized sampling, and application to the uniform resampling problem,” SIAM J. Math. Anal., vol. 45, no. 5, pp. 3132–3167, 2013. [7] W.-S. Tang, “Oblique projections, biorthogonal Riesz bases and multiwavelets in Hilbert spaces,” Proc. Amer. Math. Soc., vol. 128, no. 2, pp. 463–473, 2000. [8] K. Guo, G. Kutyniok, and D. Labate, “Sparse multidimensional representations using anisotropic dilation and shear operators,” Wavelets and splines: Athens 2005, vol. 1, pp. 189–201, 2006. [9] D. Labate, W.-Q. Lim, G. Kutyniok, and G. Weiss, “Sparse multidimensional representation using shearlets,” Wavelets XI., Proceedings of the SPIE, vol. 5914, pp. 254–262, 2005. [10] P. Kittipoom, G. Kutyniok, and W.-Q. Lim, “Construction of compactly supported shearlet frames,” Constructive Approximation, vol. 35, no. 1, pp. 21–72, 2012. [11] P. Grohs, “Continuous shearlet frames and resolution of the wavefront set,” Monatshefte für Mathematik, vol. 164, no. 4, pp. 393–426, 2011. [12] O. Christensen, An introduction to frames and Riesz bases, ser. Applied and Numerical Harmonic Analysis. Birkhäuser Boston, Inc., Boston, MA, 2003. [13] G. Kutyniok, W.-Q. Lim, and X. Zhuang, “Digital shearlet transform,” CoRR, vol. abs/1108.0454, 2011. [14] W.-Q. Lim, “Nonseparable shearlet transform,” IEEE Transactions on Image Processing, vol. 22, no. 5, pp. 2056–2065, 2013. [15] G. Kutyniok and W.-Q. Lim, “Dualizable shearlets and sparse approximation,” preprint, 2014. [16] B. Adcock, A. C. Hansen, and C. Poon, “On optimal wavelet reconstructions from Fourier samples: Linearity and universality of the stable sampling rate,” Appl. Comput. Harmon. Anal., to appear. [17] B. Adcock, A. C. Hansen, G. Kutyniok, and J. Ma, “Linear stable sampling rate: Optimality of 2D wavelet reconstructions from fourier measurements,” SIAM J. Math. Anal., to appear. [18] J. Ma, “Generalized sampling reconstruction from Fourier data using compactly supported shearlets,” preprint, 2015. [19] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “Sense: Sensitivity encoding for fast mri,” Magnetic resonance in medicine : official journal of the Society of Magnetic Resonance in Medicine / Society of Magnetic Resonance in Medicine, vol. 42, no. 5, 1999. [20] G. Kutyniok, W.-Q. Lim, and R. Reisenhofer, “ShearLab 3D: Faithful digital shearlet transforms based on compactly supported shearlets,” ACM Transactions on Mathematical Software, vol. 42, no. 1, 2015. [21] M. Lustig, D. Donoho, and J. Pauly, “Sparse MRI: The application of compressed sensing for rapid MR imaging,” Magn. Reson. Med., vol. 58, no. 6, pp. 1182–1195, 2007. [22] D. L. Donoho, “Sparse components of images and optimal atomic decomposition,” Constr. Approx., vol. 17, no. 3, pp. 353–382, 2001. [23] G. Kutyniok and D. Labate, Shearlets: Multiscale Analysis for Multivariate Data. Birkhäuser Basel, 2012.