Lecture 1 Sampling of Signals by Graham C. Goodwin University of Newcastle Australia Lecture 1 Presented at the “Zaborszky Distinguished Lecture Series” December 3rd, 4th and 5th, 2007 1 Recall Basic Idea of Sampling and Quantization 6 5 4 Quantization 3 2 1 t 0 t1 t3 t2 t4 Sampling 2 In this lecture we will ignore quantization issues and focus on the impact of different sampling patterns for scalar and multidimensional signals 3 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 4 Sampling: Assume amplitude quantization sufficiently fine to be negligible. Question: Say we are given f (ti ) ; i Î Z Under what conditions can we recover f (t ); t Î ¡ from the samples? 5 A Well Known Result (Shannon’s Reconstruction Theorem for Uniform Sampling) Consider a scalar signal f(t) consisting of ö çç- ws , ws ÷ frequency components in the range æ ÷. If 2 2ø è this signal is sampled at period D < 2p w , then the s signal can be perfectly reconstructed from the samples using: ¥ y (t ) = å k= - ¥ éæws ö ù ú sin êçç ÷ t k D ( ) ÷ ÷ êëè 2 ø ú û y [k ] æws ö çç ÷ (t - kD ) ÷ ÷ è2ø 6 Proof: Sampling produces folding Ys (w) - ws 2 ws 2 ws Low pass filter recovers original spectrum Hence Y (w) = H s (w)Ys (w) H s (w) = 1 or y (t ) = = ò ¥ - ¥ ò ¥ - ¥ s hs (s )y (t - s )d s = 0 æ- ws ö ws ÷ çç £ w£ ÷ çè 2 ø 2÷ otherwise ¥ hs (s ) å y [k ]d(t - s - k D )d s k= - ¥ ¥ = å k= - ¥ y [k ]hs (t - k D ) 7 A Simple (but surprising) Extension [Recurrent Sampling] D k = M kD where {M k } is a periodic sequence of integers; i.e., M k + N = M k Let N å Mk = K k= 1 T = KD KD =D Note that the average sampling period is e.g. N D1 = 9 D2 = 1 D3 = 9 D4 = 1 average 5 8 x x xx x x -1 0 9 10 19 20 x 0 Non-uniform x x x x 5 10 15 20 Uniform 9 Claim: Provided the signal is bandlimited to æççè- ws 2 , ws 2 öø÷÷ where ws = 2p D , then the signal can be perfectly reconstructed from the periodic sampling pattern. where D = average sampling period Proof: We will defer the proof to later when we will use it as an illustration of Generalized Sampling Expansion (GSE) Theorem. 10 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 11 Multidimensional Signals Digital Photography x1 x2 Digital Video x2 x1 x3 (time) 12 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 13 How should we define sampling for multidimensional signals? Utilize idea of Sampling Lattice T nonsingular matrix Î ¡ D ´ ¡ D Sampling Lattice L = Lat (T ) = {Tn : n Î Z D } 14 Also, need multivariable frequency domain concepts. These are captured by two ideas i. ii. Reciprocal Lattice Unit Cell 15 Reciprocal Lattice { L = Lat 2p (T * T - 1 ) } { = 2p (T T - 1 ) n:nÎ ZD } Unit Cell UC (L * ) (Non-unique) i. {UC (L )+ 2p (T * n1 , n2 Î Z D ii. U{ T - 1 ) nÎ Z T - 1 ) } n2 = Æ n1 ¹ n2 UC (L )+ 2p (T D }{ n1 Ç UC (L )+ 2p (T * T - 1 ) } n = RD 16 One Dimensional Example x x -20 -10 x 0 D x x 10 20 Sampling Lattice L = {D .n : n Î Z } 17 Reciprocal Lattice and Unit Cell Unit Cell w1 0 1 10 2 10 3 10 2p 18 Multidimensional Example x2 5 4 3 2 1 -4 -3 -2 -1 1 2 3 4 5 x1 -1 -2 -3 -4 é2 1ù ú T= ê êë0 2ú û 19 Reciprocal Lattice and Unit Cell for Example w2 2p ( - 1 UC 2p (T T ) ) 1/2 1/4 -1/4 -1/2 -3/4 -1 1/4 1/2 3/4 1 w1 2p é1 ê ê2 T - 1 (T ) = ê 1 êê ë 4 ù 0ú ú ú 1ú ú 2û 20 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 21 We will be interested here in the situation where the Sampling Lattice is not a Nyquist Lattice for the signal (i.e., the signal cannot be perfectly reconstructed from the original pattern!) Strategy: We will generate other samples by ‘filtering’ or ‘shifting’ operations on the original pattern. 22 Consider a bandlimited signal f (x), x Î ¡ D . Assume the D-dimension Fourier transform has finite support, S. Then for given D-dimensional lattice T, there always P exists a finite set {wi } Î L * , such that support ( fˆ (w) Í S = ) P 1 ( ) * UC L ( )+ wi . U i= 1 Heuristically: The idea of “Tiling” the area of interest in the frequency domain 23 One Dimensional Example Our one dimensional example continued. Sampling Lattice L = {kD ; k Î Z } Unit Cell w1 0 fˆ (w) 1 10 2 10 3 10 2p Bandlimited spectrum - 1 12 w1 = 0 Use æ2p ö Support w2 = - çç ÷ ÷ çè 10 ÷ ø w 1 12 2p ( fˆ (w)) = éêëUC (L )ùúûUéêëUC (L )+ w ùúû * * 2 24 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 25 Generation of Extra Samples Suppose now we generate a data set Q as shown in below (Q ³ P) g q (Tn) {{ }q= 1}nÎ Z D ĥ1 (w) f (x ) g1 (x) L M gq (x) hˆq (w) L M gQ (x) ˆ hQ (w) L Q – Channel Filter Bank 26 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 27 éhˆ (w + w ) hˆ (w + w ) L 1 2 1 ê1 êˆ êh1 (w + w2 ) Define H (w) = ê M ê ê êhˆ1 (w + wP ) ë Let hˆQ (w + w1 )ù ú ú ú ú ú ú ˆ hQ (w + wP )ú û éF 1 (w, x )ù ê ú F (w, x ) = êê M úú êF (w, x )ú êë Q úû ée jw1T x ù ê ú be the solution (if it exists) of H (w)F (w, x ) = êê M úú ê jwTP x ú êëe úû * for w Î UC (L ) 28 Conditions for Perfect Reconstruction GSE Theorem: f (x) can be reconstructed from Q f (x) = å å g q (Tk )f q (x - Tk ) q= 1 k Î Z D if and only if H (w) has full row rank for all w in the Unit Cell where f q (x)= òF UC q (w, x)e jwT x dw 29 Proof: F q (w, x)e jwT x = å - jwT Tk f q (x - Tk )e ; w Î UC (L * ) kÎ Z D Multiply both sides by hˆq (w + wi ) where wi Î L * (the Reciprocal Lattice). Then sum over ‘q’ Q å å q= 1 k Î Z hˆq (w + wi )f q (x - Tk )e - jwT Tk T hˆq (w + wi )F q (w, x )e jw x å = D q= 1 T =e Q j(w+ wi ) x from the Matrix identity that defines F (w, x ) Note that {w + wi ; w Î L * and i = 1,K , P} “tiles” the entire support S Thus, f (x ) = ò fˆ (w)e P dw = fˆ (w + wi )e å ò T j(w+ wi ) x dw i= 1 UC L * ( ) s P = jwT x å ò i= 1 UC L * ( ) Q fˆ (w + wi )å å q= 1 k Î Z D hˆq (w + wi )f q (x - Tk )e- jwT Tk dw 30 éP ê f (x) = å å êå ò fˆ (w + wi )hˆq (w + wi )e q= 1 k Î Z D êi= 1 UC L * ( ) êë Q ù ú j(w+ wi ) Tk d wúf q (x - Tk ) ú ú û T where we have used the fact that wi = 2p (T T - 1 ) l for l Î Z D . Since gq (x) is the output of f(x) passing through hˆq (w), then [ ]= gq (Tk ) Hence, we finally have Q f (x) = å å q= 1 k Î Z D g q (Tk )f q (x - Tk ) ÑÑÑ 31 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 32 Special Case: Recurrent Sampling (where hq is implemented by a “spatial” shift xq ) This amounts to the sampling pattern: Q Y= I {Lat (T )+ x } q q= 1 where w.l.o.g. {xq }Î UC (T ) Now, given the samples {f (x%)}x%Î Y , our goal is to perfectly reconstruct f (x). 33 jw x Here hˆq (w) = e q , and g q (x) = f (x + xq ) T Thus {g q (Tn)}nÎ Z P = ,q= 1,K Q {f (x%)}x%Î Y To apply the GSE Theorem we require T j(w+ w1 ) xQ ù é j(w+ w1 )T x1 L e êe ú H (w) = ê ú T T j w + w x ( ) j w + w x ( ) P Q ú P 1 êe L e ë û ée jw1T x1 L e jw1T xQ ùé jwT x1 ù 0 e úê ú = êê T O T ú jw xQ ú jwTP xQ ê jwP x1 0 e êëe ú ú L e û ûêë Nonsingular 34 Something to think about The GSE result depends on inversion of a particular matrix, H(w). Of course we have assumed here perfect representation of all coefficients. An interesting question is what happens when the representation is imperfect i.e. coefficients are represented with finite wordlength (i.e. they are quantized) We will not address this here but it is something to keep in mind. 35 Return to our one-dimensional example Recall that we had w1 = 0 2p w2 = 10 so that * éUC (L * )+ w ù È support fˆ (w) = éêëUC (L )ù 2ú ú û êë û ( ) Say we use recurrent sampling with x1 = 0 x2 = 0.9D ; D = 10 36 x1 = 0 x x x 0 10 20 x x x -1 0 9 19 x2 = 0.9D xx xx xx -1 0 9 10 19 20 37 Condition for Perfect Reconstruction is ée jw1x1 e jw1x2 ù ê ú nonsingular êe jw2 x1 e jw2 x2 ú ë û é1 ù 1 = ê - j(0.9)(2 p )ú ê1 e ú ë û Hence, the original signal can be recovered from the sampling pattern given in the previous slide. 38 Summary We have seen that the well known Shannon reconstruction theorem can be extended in several directions; e.g. Multidimensional signals Sampling on a lattice Recurrent sampling Given specific frequency domain distributions, these can be matched to appropriate sampling patterns. 39 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 40 Application: Video Compression Source Introduction to video cameras Instead of tape, digital cameras use 2D sensor array (CCD or CMOS) DVCD controller Image Image Processing Processor Processor Pipeline Display ( TV or LCD ) Memory 41 Image Sensor A 2D array of sensors replaces the traditional tape Each sensor records a 'point' of the continuous image The whole array records the continuous image at a particular time instant 42 2D Colours Sensor Array Data transfer from array is sequential and has a maximal rate of Q. 43 * Based on http://www.dpreview.com/learn/ Current Technology Uniform 3D sampling a sequence of identical frames equally spaced in time 44 Video Bandwidth w x depends on spatial resolution of the frames x depends on the frame rate The volume of ‘box’ depends on the capacity: pixel rate = (frame rate) x (spatial resolution) 45 Hard Constraints 1. Data recording on sensor: • Sensor array density - for spatial resolution • Sensor exposure time - for frame rate R pixels frame F frames sec. 2. Data reading from sensor: • Data readout time - for pixel rate Q pixels sec. 46 BUT... Generally Need: s.t. Q << RF R1< RF1 < F R1F1 = Q Compromise: spatial resolution temporal resolution R1< R F1 < F 47 Actual Capacity (Data Readout) wx volume determined by Q R1 F1 wy wt R1 F1 48 Observation Most energy of typical video scene is concentrated around the w x ,w y plane and the w t axis. wt wx 49 The Spectrum of this Video Clip uniform sampling - no compromise uniform sampling - compromise in frame rate uniform sampling - compromise in spatial resolution wt wx 50 Recurrent Non-Uniform Sampling 2 M 1 1 t 2 M 2 1 t t t frame type A x 2 L 1 x frame type B y 2 N 1 y 52 What Does it Buy? wx wt wy 53 Schematic Implementation non-uniform data from the sensor t Filter bank t t uniform high def. video 'compression at the source' 54 Recurrent Non-Uniform Sampling A special case of Generalized Sampling Expansion Theorem 55 Sampling Pattern The resulting sampling pattern is given by: lx 2 M1 0 LAT (U ) LAT (U ) LAT (U ) 0 m t l 1 m2 M 2 1 2L 2 ( L M ) 1 LAT (U ) x s 1 s 56 Frequency Domain S UC(2U 2 ( L M ) 1 r 1 where: T ) wr w UC (2U T ) w x : w x , wt w ( 2 L 1 ) x ( 2 M 1 ) t t 1 is the unit cell of the reciprocal lattice (2 L 1)x T LAT (2U ) 0 2 n : n Z (2M 1 1)t 0 57 Reciprocal Lattice wt 2M 1 ( 2 M 1 1) t Unit cell (2 M 1 1) t wx x x (2 L 1)x 2M 1 ( 2 M 1 1) t 58 Apply the GSE Theorem 1 2 ( x) H (w ) 2( L M )1 where: H (w) is uniquely defined by H1…H2(…) is a set of 2(L+M)+1 constraints If H 1 exists, we can find the reconstruction function (w , x) H (w ) ( x) 1 59 Reconstruction Scheme I(x,t) H1 1 H2L+1 2L+1 H2(L+M)+1 2(L+M)+1 H r .s e jw T r Î(x,t) xr tr The sub-sampled frequency of each filter H is: Nyquist frequency 2(L M) 1 60 Reconstruction functions sin sin ( x (r - 1)x) x (2M 1)t r ( x, t ) (2 L 1)xt ( x (r 1)x) t for r = 2,3,…,2L+1 sin (t (r - 2L - 1)t) sin (2L 1)x t r ( x, t ) (2M 1)xt x (t (r 2 L 1)t ) for r = 2(L+1),…,2(L+M)+1 Multidimensional ‘sinc like’ functions 61 Demo Full resolution sequence Temporal decimation Reconstructed sequence Spacial decimation 62 Outline 1. 2. 3. 4. 5. 6. 7. 8. 9. One Dimensional Sampling Multidimensional Sampling Sampling and Reciprocal Lattices Undersampled Signals Filter Banks Generalized Sampling Expansion (GSE) Recurrent Sampling Application: Video Compression at Source Conclusions 63 Conclusions Nonuniform sampling of scalar signals Nonuniform sampling of multidimensional signals Generalized sampling expansion Application to video compression A remaining problem is that of joint design of sampling schemes and quantization strategies to minimize error for a given bit rate 64 References One Dimensional Sampling A. Feuer and G.C. Goodwin, Sampling in Digital Signal Processing and Control. Birkhäuser, 1996. R.J. Marks II, Ed., Advanced Topics in Shannon Sampling and Interpolation Theory. New Your: Springer-Verlag, 1993. Multidimensional Sampling W.K. Pratt, Digital Image Processing, 3rd ed: John Wiley & Sons, 2001. B.L. Evans, “Designing commutative cascades of multidimensional upsamplers and downsamplers,” IEEE Signal Process Letters, Vol4, No.11, pp.313-316, 1997. Sampling and Reciprocal Lattices, Undersampled Signals A.Feuer, G.C. Goodwin, ‘Reconstruction of Multidimensional Bandlimited Signals for Uniform and Generalized Samples,’ IEEE Transactions on Signal Processing, Vol.53, No.11, 2005. A.K. Jain, Fundamentals of Digital Image Processing, Englewood Cliffs, NJ: PrenticeHall, 1989. 65 References Filter Banks Y.C. Eldar and A.V. Oppenheim, ‘Filterbank reconstruction of bandlimited signals from nonuniform and generalized samples,’ IEEE Transactions on Signal Processing, Vol.48, No.10, pp.2864-2875, 2000. P.P. Vaidyanathan, Multirate Systems and Filter Banks. Englewood Cliffs, NJ: Prentice-Hall, 1993. H. Bölceskei, F. Hlawatsch and H.G. Feichtinger, ‘Frame-theoretic analysis of oversampled filter banks,’ IEEE Transactions on Signal Processing, Vol.46, No.12, pp.3256-3268, 1998. M. Vetterli and J. Kovaĉević, Wavelets and Subband Coding, Englewood Cliffs, NJ: Prentice Hall, 1995. 66 References Generalized Sampling Expansions, Recurrent Sampling A. Papoulis, ‘Generalized sampling expansion,’ IEEE Transaction on Circuits and Systems, Vol.CAS-24, No.11, pp.652-654, 1977. A. Feuer, ‘On the necessity of Papoulis result for multidimensional (GSE),’ IEEE Signal Processing Letters, Vol.11, No.4, pp.420-422, 2004. K.F.Cheung, ‘A multidimensional extension of Papoulis’ generalized sampling expansion with application in minimum density sampling,’ in Advanced Topics in Shannon Sampling and Interpolation Theory, R.J. Marks II. Ed., New York: SpringerVerlag, pp.86-119, 1993. Video Compression at Source E. Shechtman, Y. Caspi and M. Irani, ‘Increasing space-time resolution in video’, European Conference on Computer Vision (ECCV), 2002. N. Maor, A. Feuer and G.C. Goodwin, ‘Compression at the source of digital video,’ To appear EURASIP Journal on Applied Signal Processing. 67 Lecture 1 Sampling of Signals by Graham C. Goodwin University of Newcastle Australia Lecture 1 Presented at the “Zaborszky Distinguished Lecture Series” December 3rd, 4th and 5th, 2007 68