Subband and wavelet coding

                 

Vector convolution, convolutional transforms Filter banks vs. vector space interpretation Orthogonal and biorthogonal subband transforms DCT as a filter bank Lapped Orthogonal Transform (LOT) Discrete Wavelet Transform (DWT) Quadrature mirror filters and conjugate quadrature filters Lifting Embedded zero-tree coding of wavelet coefficients

Bernd Girod: EE398A Image and Video Compression

Subband and Wavelet Coding no. 1

Subband coding: motivation
 

 

 

Coding with block-wise transform introduces visible blocking artifacts, as bit-rate decreases. Can we, somehow, overlap adjacent blocks,   thereby smoothing block boundaries,   but without increasing the number of transform coefficients? Solution: subband transform.

Bernd Girod: EE398A Image and Video Compression

Subband and Wavelet Coding no. 2

Vector convolution

+

+

Forward transform Inverse transform
Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 3

4 .Perfect reconstruction condition   Original domain   z-transform: “polyphase matrices”   Perfect reconstruction condition in the z-domain   Example. m=2 Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

Filter bank interpretation of convolutional transform Analysis filterbank subband signals Synthesis filterbank + Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 5 .

6 .Frequency domain perspective Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

7 .Vector space interpretation   Subband decomposition is the projection of the input onto a set of “analysis vectors” in the Hilbert space of square summable sequences “Analysis vector”
 Consider signal in channel q n denotes shift     Synthesis filterbank is linear combination of synthesis “basis vectors” “Synthesis vector”
 n denotes shift Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

Orthonormal subband transforms   Orthonormal expansion   Analysis and synthesis vectors are identical! Vector space Convolutional  transform Filter bank Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 8 .

9 . but each analysis vector must be orthogonal to all but one synthesis vectors (and vice versa)     Equivalent to perfect reconstruction Important for linear-phase FIR filters. since lapped orthogonal transforms with linear phase do not exist. Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.Biorthogonal transforms   Analysis vectors and synthesis vectors are not necessarily each orthogonal.

Subbands vs. 10 . Linear phase is desirable for images.     Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. block-wise transform   Blockwise transforms are a special case of subband decompositions with:   Number of bands m = order of transform N   Length of impulse responses of analysis/synthesis filters ≤ m Filters used in subband coders are not in general orthogonal.

block-wise transform (cont.Subbands vs. 11 .) Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

12 .Frequency response of a DCT of order N=8 10 0 -10 -20 -30 0 Frequency response [dB] Frequency response [dB] 10 Frequency response [dB] Frequency response [dB] 10 Frequency response [dB] Frequency response [dB] i=0 0 0 -10 -10 -20 -20 π/2 Frequency π -30 0 i=1 π/2 Frequency π/2 Frequency π π -30 0 i=2 10 0 -10 -20 π/2 Frequency π/2 Frequency π π -30 0 i=3 π/2 Frequency π/2 Frequency π π 10 0 -10 -20 -30 0 Frequency response [dB] Frequency response [dB] Frequency response [dB] Frequency response [dB] 10 Frequency response [dB] Frequency response [dB] 10 0 0 Frequency response [dB] Frequency response [dB] i=4 π -10 -20 π/2 Frequency π/2 Frequency π -30 0 i=5 -10 -20 π/2 Frequency π/2 Frequency π π -30 0 i=6 10 0 -10 -20 π/2 Frequency π/2 Frequency π π -30 0 i=7 π π/2 Frequency π/2 Frequency π Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

13 .Lapped Orthogonal Transform   Orthonormal convolutional transform with perfect reconstruction
   Lapped orthogonal transform (LOT): only A[0] and A[1] nonzero. hence Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

1997] Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.Example LOT basis functions. EE392C class project. Levinsky. m=8 [G. 14 .

LOT vs. 1997] Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. EE392C class project. Levinsky.453 bpp [G.426 bpp DCT quantizer step size 70 entropy 0. DCT coding LOT quantizer step size 70 entropy 0. 15 .

16 .Discrete Wavelet Transform   Recursive application of a two-band filter bank to the lowpass band of the previous stage yields octave band splitting: frequency   Same concept can be derived from wavelet theory: Discrete Wavelet Transform (DWT) Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

Cascaded analysis / synthesis filterbanks Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 17 .

2-d Discrete Wavelet Transform Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 18 .

19 .2-d Discrete Wavelet Transform example Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

20 .2-d Discrete Wavelet Transform example Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

21 .2-d Discrete Wavelet Transform example Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

2-d Discrete Wavelet Transform example Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 22 .

2-d Discrete Wavelet Transform example Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 23 .

24 .Two-channel filterbank 2 2 2 2 ∑ Aliasing   Aliasing cancellation if : Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

Example: two-channel filter bank with perfect reconstruction   Impulse responses. 25 . Tabatabai. 1988] 0 0 π 2 frequency ω π Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. analysis filters: Lowpass Highpass     Mandatory in JPEG2000 Frequency responses: frequency response h0 e jω ( ) g1 e jω ( ) 2 g0 e jω   Impulse responses. synthesis filters Lowpass Highpass 1 ( ) h1 e jω ( ) “Biorthogonal 5/3 filters” “LeGall-Tabatabai filters” [LeGall.

Galand. 1976]     Highpass band is the mirror image of the lowpass band in the frequency domain Need to design only one prototype filter Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.Quadrature Mirror Filters (QMF)   QMFs achieve aliasing cancellation by choosing Example: 16-tap QMF filterbank [Croisier. Esteban. 26 .

Conjugate quadrature filters   Achieve aliasing cancelation by Prototype filter [Smith. 27 . 1986]   Impulse responses   Perfect reconstruction: find power complementary prototype filter   Orthonormal subband transform! Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. Barnwell.

Lifting   Analysis filters Σ Σ K0 Σ Σ K1     L “lifting steps” [Sweldens 1996] First step can be interpreted as prediction of odd samples from the even samples Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. 28 .

Lifting (cont. 29 .)   Synthesis filters Σ - Σ - - Σ - Σ     Perfect reconstruction (biorthogonality) is directly built into lifting structure Powerful for both implementation and filter/wavelet design Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

30 .Example: lifting implementation of 5/3 filters Σ 1+ z −1 4 Σ 1/2 Verify by considering response to unit impulse in even and odd input channel. Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

Reversible subband transform     Observation: lifting operators can be nonlinear. 31 . Incorporate the necessary rounding into lifting operator: Σ Σ K0 Σ   Σ K1 Used in JPEG2000 as part of 5/3 biorthogonal wavelet transform Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.

074 bpp 0. 32 . 2001] Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.048 bpp Original 512x512 8bpp Error images enlarged [Gonzalez.Wavelet compression results 0. Woods.

but some descendants are significant POS: positive significant NEG: negative significant   For the highest bands.  X X X X X X X X X X X X XXXX X X X X X X X X X X X X XXXX 3. children) Coefficient magnitude > threshold: significant coefficients Four cases 1. coefficient and all descendants are not significant IZ: coefficient is not significant. 33 .  XX XX 2.  4.  ZTR: zero-tree. ZTR and IZ symbols are merged into one symbol Z Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no.Embedded zero-tree wavelet algorithm X XX XX X X X X X X X X X X X X XXXX   X X „Parent“ XX XX   „Children“   Idea: Conditional coding of all descendants (incl.

)   Successive approximation quantization and encoding   Initial „dominant“ pass •  Set initial threshold θ. determine significant coefficients •  Arithmetic coding of symbols ZTR. until bit budget is exhausted. POS. 34 .   Repeat dominant pass •  Omit previously found significant coefficients •  Decrease threshold θ by factor of 2. NEG   Subordinate pass •  Refine magnitude of all coefficients found significant so far by one bit (subdivide magnitude bin by two) •  Arithmetic coding of sequence of zeros and ones. IZ. Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. IZ. POS.Embedded zero-tree wavelet algorithm (cont. NEG   Repeat subordinate and dominant passes. determine new significant coefficients •  Arithmetic coding of symbols ZTR.

Pearlman. pp. December 1993. vol. W. 41. „A new.)       Decoding: bitstream can be truncated to yield a coarser approximation: „embedded“ representation Further details: J.“ IEEE Transactions on Circuits and Systems for Video Technology. Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. A.Embedded zero-tree wavelet algorithm (cont. fast.. 3445-3462. Enhancement SPIHT coder: A. 35 . and efficient image codec based on set partitioning in hierarchical trees. vol. M. „Embedded image coding using zerotrees of wavelet coefficients. Said. 243-250. Shapiro. no. pp. 63 . June 1996.“ IEEE Transactions on Signal Processing. A. 12.

e. lossless) wavelet transform Zero-trees: exploit statistical dependencies across subbands Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no..Conclusions:
 subband and wavelet coding                 Overlapping basis functions can reduce blocking artifacts Biorthogonal subband transforms = perfect reconstruction DCT can be understood as a (poor) filter bank Discrete Wavelet Transform = cascaded dyadic subband splits Quadrature mirror filters and conjugate quadrature filters: aliasing cancellation Lifting: powerful for implementation and wavelet construction Lifting allows reversible (i. 36 .

D. April 1989.1-6.” vol. J. pp. pp. H. 7 H. Sections 4. and Signal Processing. vol. Malvar. S. pp. 2.Reading         Taubman. Le Gall. 12. Speech. Marcellin.” IEEE Transactions on Acoustics. no.” Proc. Tabatabai. “Sub-band coding of digital images using symmetric short kernel filters and arithmetic coding techniques. M. A. 1993.” IEEE Transactions on Signal Processing. 41. 761-764. Bernd Girod: EE398A Image and Video Compression Subband and Wavelet Coding no. Staelin. 37 . ICASSP-88. D. 553-559. “The LOT: transform coding without blocking effects. no. 37. vol. Shapiro.4.2. 6. “Embedded image coding using zerotrees of wavelet coefficients. 4. April 1988. 3445-3462. Dec.