Mathematics and Computation in Imaging Science and Information Processing July-December, 2003 • Organized by Institute of Mathematical Sciences and Center for Wavelet. Approximation, and Information Processing, National University of Singapore. • Collaboration with the Wavelet Center for Ideal Data Representation. • Co-chairmen of the organizing committee: • Amos Ron (UW-Madison), • Zuowei Shen (NUS), • Chi-Wang Shu (Brown University) Conferences • Wavelet Theory and Applications: New Directions and Challenges, 14 - 18 July 2003 • Numerical Methods in Imaging Science and Information Processing, 15 -19 December 2003 Confirmed Plenary Speakers for Wavelet Conference • • • • • • Albert Cohen Wolfgang Dahmen Ingrid Daubechies Ronald DeVore David Donoho Rong-Qing Jia • • • • • Yannis Kevrekidis Amos Ron Peter Schröder Gilbert Strang Martin Vetterli Workshops • IMS-IDR-CWAIP Joint Workshop on Data Representation, Part I on 9 – 11, II on 22 - 24 July 2003 • Functional and harmonic analyses of wavelets and frames, 28 July - 1 Aug 2003 • Information processing for medical images, 8 - 10 September 2003 • Time-frequency analysis and applications, 22- 26 September 2003 • Mathematics in image processing, 8 - 9 December 2003 • Industrial signal processing (TBA) • Digital watermarking (TBA) Tutorials • A series of tutorial sessions covering various topics in approximation and wavelet theory, computational mathematics, and their applications in image, signal and information processing. • Each tutorial session consists of four one-hour talks designed to suit a wide range of audience of different interests. • The tutorial sessions are part of the activities of the conference or workshop associated with. Membership Applications • To stay in the program longer than two weeks • Please visit http://www.ims.nus.edu.sg for more information Wavelet Algorithms for High-Resolution Image Reconstruction Zuowei Shen Department of Mathematics National University of Singapore http://www.math.nus.edu.sg/~matzuows Joint work with (accepted by SISC) T. Chan (UCLA), R.Chan (CUHK) and L.X. Shen (WVU) Outline of the talk Part I: Problem Setting Part II: Wavelet Algorithms What is an image? image = matrix pixel intensity = matrix entry Resolution = size of the matrix I. High-Resolution Image Reconstruction: Resolution = 64 64 Resolution = 256 256 Four low resolution images (64 64) of the same scene. Each shifted by sub-pixel length. Construct a high-resolution image (256 256) from them. Boo and Bose (IJIST, 97): #4 #2 taking lens #1 partially silvered mirrors relay lenses CCD sensor array Four 2 2 images merged into one 4 4 image: a1 a2 b1 b2 a3 a4 b3 b4 a1 b1 a2 b2 c1 d1 c2 d2 a3 b3 a4 b4 c1 c2 d1 d2 c3 d3 c4 d4 d4 Observed highresolution image c3 c4 d3 Four low resolution images By permutation Four 64 64 images merged into one by permutation: Observed highresolution image by permutation Modeling Consider: Low-resolution pixel 1 4 1 2 1 4 1 2 1 1 2 1 4 1 2 1 4 High-resolution pixels Observed image: HR image passing through a low-pass filter a. LR image: the down samples of observed image at different sub-pixel position. After modeling and adding boundary condition, it can be reduced to : L f = g , Where L is blurring matrix, g is the observed image and f is the original image. The problem L f = g is ill-conditioned. Regularization is required: ( L* L R)f L*g. Here R can be I, . It is called Tikhonov method ( or the least square ) g ( L* L) 1 L*g ( L* L R) 1 L*g Wavelet Method • Let â be the symbol of the low-pass filter. Assume: d ˆ ˆd ˆ a , b, b can be found such that • aˆ d aˆ ˆ d bˆ 1 b Z 22 \{ 0} • One can use unitary extension principle to obtain a set of tight frame systems. Let be the refinable function with refinement mask a, i.e. 4 a( ) (2 ) . Z 2 Let d be the dual function of : , d 0 . We can express the true image as f 2 v d 2 , Z 2 where v() are the pixel values of the high-resolution picture. The pixel values of the observed image are given by a * v , Z2 The observed function is g d ( a ) ( / 2) . Z 2 The problem is to find v( ) from (a * v)(). From 4 sets low resolution pixel values reconstruct f, lift 1 level up. Similarly, one can have 2 level up from 16 set... Do it in the Fourier domain. Note that aˆ d aˆ bˆd bˆ 1 . Z 22 \{ 0} We have aˆ aˆvˆ bˆd bˆ Z 2 \0 2 d vˆ vˆ . or aˆ a * v bˆd bˆ Z 2 \0 2 d vˆ vˆ . (1) Generic Wavelet Algorithm: (i) Choose vˆ0 L2 , ; 2 (ii) Iterate until convergence: vˆn 1 aˆ a * v bˆd bˆ Z 2 \0 2 d vˆn . Proposition Suppose that 0 a ˆda ˆ 1and nonzero almost everywhere. Then || vˆn vˆ || 2 0for arbitrary v̂0 . Regularization: Damp the high-frequency components in the current iterant. Wavelet Algorithm I: (i) Choose ˆv0 L2 , 2 ; (ii) Iterate until convergence: d vˆn 1 aˆ a * v (1 ) bˆ bˆ vˆn . Z 2 \0 2 d Matrix Formulation: The Wavelet Algorithm I is the stationary iteration for ( Ld L H d H )f Ld g . Different between Tikhonov and Wavelet Models: • Ld instead of L*. • Wavelet regularization operator. Both penalize high-frequency components uniformly by . Wavelet Thresholding Denoising Method: Decompose the n-th iterate, i.e. bˆ vˆn , into different scales: ( This gives a wavelet packet decomposition of n-th iterate.) J 1 bˆ vˆ n aˆ d aˆ bˆ vˆn aˆ d J Before reconstruction, j 0 j ˆ d ˆ ˆ b b aˆ b vˆn , Z 22 \0, 0 j • Denoise these coefficients bˆ aˆ bˆ vˆn of the wavelet packet by thresholding method. Wavelet Algorithm II: (i) Choose vˆ0 L2 , ; 2 (ii) Iterate until convergence: ˆvn1 aˆ d a * v Z 22 \ 0 ,0 ˆb d T bˆ vˆ n Where T is a wavelet thresholding processing . 4 4 sensor array: Original LR Frame Tikhonov Algorithm I Observed HR Algorithm II 4 4 sensor array: Tikhonov Algorithm II Numerical Examples: 22 sensor array: 1 level up SNR (dB) 30 40 Tikhonov Algorithm I Algorithm II PSNR RE PSNR RE PSNR RE Iter. 32.55 0.0437 33.82 0.0377 34.48 0.0350 9 33.88 0.0375 34.80 0.0337 35.23 0.0321 12 44 sensor array: 2 level up SNR (dB) 30 40 Tikhonov Algorithm I Algorithm II PSNR RE PSNR RE PSNR RE Iter. 29.49 0.0621 29.70 0.0601 30.11 0.0579 30 30.17 0.0573 30.30 0.0566 30.56 0.0549 45 1-D Example: Signal from Donoho’s Wavelet Toolbox. Blurred by 1-D filter. Original Signal Tikhonov Observed HR Signal Algorithm II Calibration Error: High-resolution pixels Displacement error Ideal low-resolution pixel position Displaced lowresolution pixel ex Problem no longer spatially invariant. The lower pass filter is perturbed The wavelet algorithms can be modified Reconstruction for 4 4 Sensors: (2 level up) Original Tikhonov LR Frame Observed HR Wavelets Reconstruction for 4 4 Sensors: (2 level up) Tikhonov Wavelets Numerical Results: 2 2 sensor array (1 level up) with calibration errors: Least Squares Model Our Algorithm SNR(dB) PSNR RE * PSNR RE Iterations 30 28.00 0.0734 0.0367 30.94 0.0524 2 40 28.24 0.0715 0.0353 31.16 0.0511 2 4 4 sensor array (2 level) with calibration errors: Least Squares Model Our Algorithm SNR(dB) PSNR RE * PSNR RE Iterations 30 24.63 0.1084 0.0492 27.80 0.0752 5 40 24.67 0.1078 0.0505 26.81 0.0751 6 Super-resolution: not enough frames Example: 4 4 sensor with missing frames: (0,0) (0,1) (0,2) (0,3) (1,0) (1,1) (1,2) (1,3) (2,0) (2,1) (2,2) (2,3) (3,0) (3,1) (3,2) (3,3) Super-resolution: not enough frames Example: 4 4 sensor with missing frames: (0,1) (1,0) (0,3) (1,2) (2,1) (3,0) (2,3) (3,2) Super-Resolution: Not enough low-resolution frames. i. Apply an interpolatory subdivision scheme to obtain the missing frames. ii. Generate the observed high-resolution image w. iii. Solve for the high-resolution image u. iv. From u, generate the missing low-resolution frames. v. Then generate a new observed high-resolution image g. vi. Solve for the final high-resolution image f. Tikhonov Algorithm I Algorithm II PSNR RE PSNR RE PSNR RE 27.44 0.0787 27.82 0.0753 27.76 0.0758 Reconstructed Image: Observed LR Final Solution