The Lifting Scheme: a custom-design construction of biorthogonal wavelets Sweldens95, Sweldens 98 (appeared in SIAM Journal on Mathematical Analysis) Relations of Biorthogonal Filters ~ ( n) h ( m 2 n ) h ( m) 2 m ( n) ~ g (m 2n) g (m) 2 m h(m 2n) g~(m) 0 m ~ h ( m 2n ) g ( m) 0 m Biorthogonal Scaling Functions and Wavelets ~ Dual (t ), (t k ) (k ) Dual (t ),~(t k ) (k ) ~ wavelet dual scaling fns (t ), (t n) 0 dual wavelet scaling fns ~(t ), (t n) 0 Wavelet Transform (in operator notation) transpose ~ j H j j 1 ~ j G j j 1 Filter operators are matrices encoded with filter coefficients with proper dimensions j 1 H j G j * j * j Note that up/down-sampling is absorbed into the filter operators Operator Notation Relations on Filter Operators Biorthogonality Write in matrix form: ~ * ~ * H j H j G jG j 1 ~ * ~ * H jG j G j H j 0 ~ H j * ~ Hj G j H Exact Reconstruction * ~ *~ H j H j G jG j 1 ~ ~ j 1 H *j H j j 1 G *j G j j 1 ~ ~ H *j H j G *j G j j 1 * j 1 0 G 0 1 ~ * Hj Gj ~ 1 G j * j Theorem 8 (Lifting) • Take an initial set of biorthogonal filter operators H old j ~ old old ~ old , H j ,Gj ,Gj • A new set of biorthogonal filter operators can be found as H j , H~ j , G j , G~ j ~ • Scaling functions and H and G untouched Hj H ~ old ~ ~ old H j H j S jG j old j * old G j G old S H j j j ~ ~ old Gj Gj ~ ~ old H j 1 S H j ~ ~ old G j 0 1 G j H j 1 0 H old j old G * j S 1 G j Proof of Biorthogonality H ~ * Hj * old* G j ~ H old G j j G j * j H ~ H j * 1 * ~ H j Gj G j 0 1 0 old* j G old* j 1 0 1 0 ~ old S 1 S H j ~ old 1 0 1 G j ~ old 0 H j ~ old 1 1 G j ~ old S H j old* S old* 1 Gj ~ old H j 1 G j 0 1 S 1 0 1 S 1 0 1 0 1 0 1 0 1 Choice of S • Choose S to increase the number of vanishing moments of the wavelets • Or, choose S so that the wavelet resembles a particular shape – This has important applications in automated target recognition and medical imaging Same thing expressed in frequency domain Corollary 6. • Take an initial set of finite biorthogonal filters h, h~ 0 , g 0 , g~ • Then a new set of finite biorthogonal filters can be found as h, h~, g , g~ ~ ~ h (w ) h 0 (w ) g~ (w ) s (2w ) g (w ) g 0 (w ) h(w ) s(2w ) • where s(w) is a trigonometric polynomial Details Theorem 7 (Lifting scheme) • Take an initial set of biorthogonal scaling functions and wavelets , ~ 0 , 0 ,~ • Then a new set , ,~, ,~ which is formally biorthognal can be found as Same thing expressed in indexed notation • where the coefficients sk can be freely chosen. Dual Lifting ~ • Now leave dual scaling function and H and G filters untouched Hj H ~ ~ old Hj Hj old j ~ old S jG j G j G old j ~ ~ old ~ * ~ old Gj Gj S j H j Fast Lifted Wavelet Transform • Basic Idea: never explicitly form the new filters, but only work with the old filter, which can be trivial, and the S filter. Before Lifting ~ old j H j j 1 ~ old j G j j 1 Forward Transform ~ old ~ ~ old j H j j 1 H j S j G j j 1 ~ old H j j 1 S j j Inverse Transform j 1 H *j j G *j j H old * j j G H old * j j old * j S jH S j j G old * j j old * j j Examples Interpolating Wavelet Transform Biorthogonal Haar Transform The Lazy Wavelet • Subsampling operators E (even) and D (odd) E * D E E H * lazy j 1 0 D 0 1 * E D 1 D * ~ lazy ~ lazy lazy H j E and G j G j D Interpolating Scaling Functions and Wavelets • Interpolating filter: always pass through the data points • Can always take Dirac function as a formal dual ~ H E S jD ~ int Hj E int j G D ~ int ~* Gj D S j E int j Theorem 15 • The set of filters resulting from interpolating scaling functions, and Diracs as their formal dual, can be seen as a dual lifting of the Lazy wavelet. ~ H j H E S jD ~* ~ ~ int ~ ~ int H j H j S j G j (1 S j S j ) E S j D int * int * *~ G j G j S j H j S j E (1 S j S j ) D ~ ~ int ~* Gj Gj D S j E int j Algorithm of Interpolating Wavelet Transform (indexed form) Example: Improved Haar • Increase vanishing moments of the wavelets from 1 to 2 • We have ~0 h (w ) h(w ) 12 12 e iw 0 ~ 1 1 iw g (w ) g (w ) e 2 2 After lifting : g (w ) g 0 (w ) h(w ) s(2w ) Details Verify Biorthogonality ~ ( n) h ( m 2 n ) h ( m ) 2 m ( n) ~ g (m 2n) g (m) 2 m h(m 2n) g~(m) 0 m ~ h ( m 2n ) g ( m) 0 m ~0 hn hn { 12 12 }n 0,1 g~n g n0 { 21 12 }n 0,1 Improved Haar (cont) 0th moment van ishes : g (0) 0 g 0 (0) h(0) s(0) 1 2 12 12 s(0) 0 s(0) 0 1st moment van ishes : g ' (0) 0 g ' (w ) g 0' (w ) h' (w ) s (2w ) 2h(w ) s' (2w ) g 0' (w ) 2i e iw g ' (0) g 0 ' (0) h' (0) s (0) 2h(0) s ' (0) 2i 0 2( 12 12 ) s ' (0) 2i 2s' (0) 0 s' (0) 4i Choose : s(w ) i 4 sin w 1 e iw e iw 4 2 e iw e iw 8 ei 2w e i 2w ei 2w e i 2w s(2w ) and s(2w ) 8 8 g (w ) g 0 (w ) h(w ) s(2w ) g(0) = g’(0) =0 i 2w i 2w e e 21 12 e iw 12 12 e iw 8 161 ei 2w 161 eiw 12 12 e iw 161 e i 2w 161 e i 3w ~ ~0 h (w ) h (w ) g~ (w ) s (2w ) i 2w i 2w e e 12 12 e iw -21 12 e iw 8 161 e i 2w 161 e iw 12 12 e iw 161 e i 2w 161 ei 3w Details Verify Biorthogonality ~ ( n) h ( m 2 n ) h ( m ) 2 m ~(m) 0 h ( m 2 n ) g m ~ h ( m 2n ) g ( m) 0 ( n) ~ g (m 2n) g (m) 2 m hn { 12 12 }n 0,1 ~ 1 1 1 hn 16 16 2 g n 161 g~n { 21 1 16 1 2 1 2 n 0 ,1 } m 1 2 1 16 1 2 1 16 1 16 n 2 , 1, 0 ,1, 2 , 3 1 16 n 2 , 1, 0 ,1, 2 , 3