Adaptive Compressive Imaging Using Sparse Hierarchical Learned Dictionaries

Transcription

1 Adaptive Compressive Imaging Using Sparse Hierarchical Learned Dictionaries Jarvis Haupt University of Minnesota Department of Electrical and Computer Engineering Supported by

2 Motivation New Agile Sensing Platforms

3 A Host of New Agile Imaging Sensors Original 20% 40% samples samples

4 Overview of This Talk Fusing Adaptive Sensing and Structured Sparsity in theory and in practice

5 Background Sparse Inference and Adaptive Sensing

6 A Model for Sparsity Objects of interest are vectors x 2 R n Signal Support: S, {i : x i =0} Sparse S = k n number of nonzero signal components

7 A Sparse Inference Task Noisy Linear Observation Model: y = x + w R m n w N (0,I m m ) Support Recovery Goal: Obtain an (accurate) estimate b S = b S(y, ) of true support S

8 A Sparse Inference Task Noisy Linear Observation Model: y = x + w R m n w N (0,I m m ) Assume Sensing energy 2 F fixed: 2 F = R x i µ for all i S What conditions are necessary/sufficient for exact support recovery? (eg., such that P (S = b S) 0 as n )

9 Exact Support Recovery? Point sampling y = x + w (Sensing energy R = n) Necessary & Sufficient for Exact Support Recovery: r n µ const. log n R Uncompressed Sensing (Donoho & Jin 2004; JH, Castro, & Nowak 2010) Compressed Sensing (Genovese, Jin, & Wasserman 2009; Aeron, Saligrama & Zhao, 2010)

10 Conditions for Exact Support Recovery Uncompressed / compressed r n µ const. log n R (N + S) Uncompressed Sensing (Donoho & Jin 2004; JH, Castro, & Nowak 2010) Compressed Sensing (Genovese, Jin, & Wasserman 2009; Aeron, Saligrama & Zhao, 2010) Question: Can we do better by exploiting structure, or adaptivity, or both?

11 Conditions for Exact Support Recovery r n µ const. log n R r n µ const. log k R Uncompressed/ Compressed (N + S) Necessity: (Castro 2012) Sufficiency (uncompressed): (Malloy & Nowak, 2010; Malloy & Nowak, 2011) Sufficiency (compressed): (JH, Baraniuk, Castro, & Nowak 2012, Malloy & Nowak 2013) y = x + w R m n w N (0,I m m ) k k 2 F = R

12 Beyond Simple Sparsity The Role of Structure

13 Our Focus: Tree Sparsity Tree T with n nodes and degree d Characteristics of tree structure: Elements of x in one-to-one correspondence with nodes of T Nonzeros of tree-sparse vector form rooted connected subtree of T Question: Does tree structure help in support recovery?

14 Conditions for Exact Support Recovery Detection of simple trail (uncompressed sensing) r n µ const. log n R r n µ const. R (N+S)* r n µ const. log k R Signal Detection Problem: (Arias-Castro, Candes, Helgason, & Zeitouni 2008)

15 Conditions for Exact Support Recovery The intersection of adaptivity and (tree) structure... r n µ const. log n R r n µ const. R r n µ const. log k R s k µ const. log k R (S) (A. Soni & JH, 2011) Akshay Soni University of Minnesota Recent related work: Adaptivity and Structure in finding activated blocks in a matrix (Balakrishnan, Kolar, Rinaldo, and Singh 2012)

16 Adaptive Tree Sensing: An Example If the hypothesis test is correct at each step, then m = dk +1=O(k) y (1) = x 1 + w (1) suppose y (1) > Adaptive Wavelet Tree Sensing in the Literature: Non-Fourier encoded MRI (Panych & Jolesz, 1994) Compressive Imaging (Deutsch, Averbuch, & Dekel, 2009) y (2) = x 2 + w (2) suppose y (2) < (none analyzed the case of noisy measurements...) y (3) = x 5 + w (3) y (4) = x 6 + w (4) y (5) = x 7 + w (5) suppose y (3) > suppose y (4) < suppose y (5) <

17 Orthogonal Dictionaries and Tree Sparsity Consider signals z R p that are sparse in a known dictionary D R p n. That is, z = Dx, where x R n is k-sparse, D satisfies D T D = I n n, and columns of D are d j, j =1, 2,...,n We are interested in the case where x is tree-sparse... Collect (noisy) observations of z by projecting onto (scaled) columns of D. Suppose, for example, that the j-th measurement is obtained by projecting onto columnn d i, then y (j) = d T i z + w (j) where w (j) N (0, 1). Nonnegative scaling factor (equivalently, could consider non-unit noise variance)

18 Support Recovery via Adaptive Tree Sensing Theorem (A. Soni & JH, 2011) Let T n,d be a balanced, rooted connected tree of degree d with n nodes. Suppose that z R p can be expressed as z = Dx, where D is a known dictionary with orthonormal columns and x is k-sparse. Further, suppose the support of x corresponds to a rooted connected subtree of T n,d. Observations of z are of the form of projections of z onto columns of D. Let the index corresponding to the root of T n,d be known, and apply the top-down tree sensing procedure with threshold and scaling parameter. For any c 1 > 0 and c 2 (0, 1), there exists a constant c 3 > 0 such that if µ =min i S x i p c 3 2 log k and = c 2 µ, the tree sensing procedure collects m = dk+1 measurements, and produces a support estimate b S that equals S with probability at least 1 k c 1. q Choose = R (d+1)k, then the theorem guarantees exact support recovery (whp) when µ s c 3 (d + 1) k log k R

19 Conditions for Exact Support Recovery The intersection of adaptivity and (tree) structure... r n µ const. log n R r n µ const. R (*conjecture for support recovery) r n µ const. log k R s k µ const. log k R (S) (A. Soni & JH, 2011) Akshay Soni University of Minnesota

20 LASeR Learning Adaptive Sensing Representations

21 Beyond Wavelet Trees: Learned Representations Given training data Z R p q, want to learn a dictionary D so that Z DX, D R p n, X R n q, and each column of X, x i R n, is tree-sparse in T n,d. Pose this as an optimization: {D, X} = arg min D R p n,d T D=I n n, {x i } qx z i Dx i (x i ) i=1 The regularization term is (x i )= P g G g (x i ) g, where G denotes a set of (overlapping) groups of indices for x, (x i ) g is x i restricted to the indices in the group g G, g are non-negative weights, and the norm can be, eg., 2 or Solve by alternating minimization over D and X (Jenatton, Mairal, Obozinski, & Bach, 2010) Sparse Modeling Software (SPAMS):

22 Group Specifications to Enforce Tree Structure Example: Binary Tree, 15 nodes, 4 levels Number of groups same as number of nodes (but varying sizes)

23 LASeR An Illustrative Example

24 Learning Adaptive Sensing Representations Training' Data' Structured' Sparsity' LASeR' Adap3ve' Sensing' Learn representation for 163 images from Psychological Image Collection at Stirling (PICS) LASeR:'Learning'Adap3ve'Sensing'Representa3ons' Example images ( )

25 Learned Orthogonal Tree-Basis Elements (First four levels of 7 total)

26 Original Image Tree Elements Present in Sparse Representation

27 Qualitative Results Sensing Energy R = Wavelet Tree Sensing m = 20 m = 50 m = 80 PCA CS LASSO CS Tree LASSO LASeR

28 Qualitative Results Sensing Energy R = Wavelet Tree Sensing m = 20 m = 50 m = 80 PCA CS LASSO CS Tree LASSO LASeR

29 Quantitative Results Reconstruction SNR(dB) #projections Reconstruction SNR(dB) #projections Reconstruction SNR(dB) #projections

30 LASeR Imaging via Patch-wise Sensing

31 Patch-wise Sensing Experiment Motivated by EO Imaging Application (Thanks: Bob Lockheed Martin) Training Data: 3 Sample images from the Columbus Large Image Format (CLIF) 2007 Dataset Each image is 1024x1024 Randomly extracted x32 patches (at random locations) and vectorized them into length 1024 vectors Applied PCA and LASeR (7-level 127 node binary tree) to this training data

32 Compare: PCA Basis Elements In the tree-sensing context, can view PCA sensing approach in terms of a tree of degree 1

33 Learned Orthogonal Tree-Basis Elements

34 Example: Approximation by Patch-wise Sensing Test Image (another image from CLIF database) Sense & reconstruct non-overlapping 32x32 patches comparing LASeR, PCA, Wavelets

35 Approximation Results Uniform Sampling Rate Sampling rate: 12.5% LASeR rsnr = 16.5 db PCA rsnr = 17.6 db rsnr, 20 log 10 (kbx xk F /kxk F )

36 Approximation Results Uniform Sampling Rate Sampling rate: 12.5% LASeR rsnr = 16.5 db rsnr, 20 log 10 (kbx xk F /kxk F )

37 Approximation Results Uniform Sampling Rate Sampling rate: 12.5% PCA rsnr = 17.6 db rsnr, 20 log 10 (kbx xk F /kxk F )

38 Approximation Results Uniform Sampling Rate Sampling rate: 12.5% LASeR rsnr = 16.5 db 2D Haar Wavelet rsnr = 13.5 db rsnr, 20 log 10 (kbx xk F /kxk F )

39 Approximation Results Adaptive Sampling Rate Average sampling rate: 7.2% LASeR rsnr = 13.9 db PCA rsnr = 15.0 db rsnr, 20 log 10 (kbx xk F /kxk F )

40 Approximation Results Adaptive Sampling Rate Average sampling rate: 7.2% LASeR rsnr = 13.9 db 2D Haar Wavelet rsnr = 11.9 db rsnr, 20 log 10 (kbx xk F /kxk F )

41 Approximation: Zoomed In A Closer Look... (Average sampling rate: 7.2%) Original LASeR PCA 2D Haar

42 Approximation: Zoomed In A Closer Look... (Average sampling rate: 7.2%) Original LASeR PCA 2D Haar

43 Sampling Rate Adapts to Block Complexity Sampling rate per block (Average sampling rate: 7.2%) LASeR PCA

44 Sampling Rate Adapts to Block Complexity Sampling rate per block (Average sampling rate: 7.2%) LASeR 2D Haar

45 Summary: Adaptivity + Structure In Theory (Support Recovery) r n µ const. log n R r n µ const. R (conj.) r n µ const. log k R s k µ const. log k R Conclusions: Polynomial reduction in SNR required for exact support recovery (for fixed sensing energy )

46 Summary: Adaptivity + Structure In Practice (Learned Representations and Patch-wise Sensing) Original LASeR PCA Conclusions: PCA works very well on small patch sizes (shared, elemental structure) in noise free settings

47 Summary: Adaptivity + Structure m = 20 m = 50 m = 80 Wavelet Tree Sensing PCA Original Image LASeR Conclusions: Potential benefit for learned representations depend on patch size, data regularity, noise. Use LASeR on PCA residuals? (ie, fuse PCA and LASeR) jdhaupt jdhaupt@umn.edu