-
Notifications
You must be signed in to change notification settings - Fork 5
/
fbank.m
239 lines (217 loc) · 7.43 KB
/
fbank.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
function [xout,a,d,err,PA,PS]=fbank(x,tipe,N,M,pph)
% [xout,a,d,err,PA,PS]=fbank(x,tipe,N,M,pph)
%
% Combines ABANK and SBANK so the perfect reconstruction
% condition becomes immediately testable.
%
% 'pph' 1= Time-domain full bitrate (inefficient);
% 2= Time-domain polyphase (inefficient);
% 3= Z-domain polyphase (default)
%
% Returns the Frequency-domain (pph=3)
% or Time domain (pph=3) polyphase matrix as
% PA (P-analysis) and PS (P-synthesis)
%
% Single-iteration filter-bank of signal 'x' with filters
% of a certain 'name'.
%
% Returns reconstructed signal without delay or distortion,
% and the rms error of the approximation.
%
% See also WTLIFT.
% Last modified by fjsimons-at-alum.mit.edu, 03/01/2002
t0=clock;
defval('pph',3)
[h0,f0]=wc(tipe,[N M]);
% Check normalization of the product filter and
% create synthesis coefficients from the analysis ones:
% conjugate mirror filters with alternating signs to satisfy
% the no-alias condition
[h0,h1,f0,f1,l,lf0,lh0]=prodco(f0,h0);
% Get downsampling arrays based on the data length
% and the filter lengths.
[d2fx,d2hx,lx,lxout]=landd(f0,h0,x);
% Get polyphase matrices as cell arrays and in components
[PA,PS,I,...
H0even,H0odd,H1even,H1odd,...
F0even,F0odd,F1even,F1odd]=polyphase(f0,h0);
% Analysis
switch pph
case 3
% This is the "Standard Algorithm"
disp('Z-domain polyphase implementation')
Xeven=x(even(x));
Xodd=x(~even(x));
% This here is the polyphase matrix in the z-form (not Toeplitz):
% Our definition is the transpose of Daubechies'
% It's a block matrix with Laurent polynomial coefficients
% Always need to offset them, but, depending on the length of the
% input series, may need to add an extra zero or not.
% If h0 and f0 are both odd, get different numbers of evens
% and odds.
% We're taking care of the case were lf0 and lh0 are both
% even and thus split into equal lengh parts
lh0lxe=length(H0even)+length(Xeven);
lh0lxo=length(H0odd)+length(Xodd);
lh1lxe=length(H1even)+length(Xeven);
lh1lxo=length(H1odd)+length(Xodd);
% This is tricky business, but it works
% -> for even- and odd length signals
% -> for even- and odd length filters
a=repmat(0,max(lh0lxe,lh0lxo-1),1);
d=repmat(0,max(lh1lxe,lh1lxo-1),1);
a(2:lh0lxe)=conv(H0even(:),Xeven(:));
a(1:lh0lxo-1)=a(1:lh0lxo-1)+conv(H0odd(:),Xodd(:));
d(2:lh1lxe)=conv(H1even(:),Xeven(:));
d(1:lh1lxo-1)=d(1:lh1lxo-1)+conv(H1odd(:),Xodd(:));
case 2
disp('Time-domain polyphase implementation')
% Filter Toeplitz matrices
% Size of H0 is (lx+lh0-1) by (lx)
% Size of H1 is (lx+lf0-1) by (lx)
% and element (1,1) is h0(1)
H0=convmtx(h0(:),lx);
H1=convmtx(h1(:),lx);
% Downsample: delete rows: keep these
AL=H0(d2hx,:);
AB=H1(d2fx,:);
% Polyphase form
% Split signal in even and odd
% Note that this is dependent on your first sample:
% do you call it x(0) or x(1)
% The Matlab convention differs from the others,
% so our even is their odd.
X=[x(even(x)) ; x(~even(x))];
% Construct the polyphase ANALYSIS matrix (Toeplitz)
H0even=AL(:,even(x));
H0odd=AL(:,~even(x));
H1even=AB(:,even(x));
H1odd=AB(:,~even(x));
% Implement filters by multiplication
% This is the polyphase applied to the samples in the time domain
PA=[H0even H0odd ;
H1even H1odd];
ad=PA*X;
% This is the intermediate vector; split up if you want
a=ad(1:sum(d2hx));
d=ad(sum(d2hx)+1:end);
case 1
disp('Toeplitz implementation')
% Filter Toeplitz matrices
% Size of H0 is (lx+lh0-1) by (lx)
% Size of H1 is (lx+lf0-1) by (lx)
% and element (1,1) is h0(1)
H0=convmtx(h0(:),lx);
H1=convmtx(h1(:),lx);
% Downsample: delete rows: keep these
AL=H0(d2hx,:);
AB=H1(d2fx,:);
% Apply - this yields the decomposition we're after
a=AL*x(:);
d=AB*x(:);
% And they have length ceil((lx+lh0-1)/2) and ceil((lx+lf0-1)/2)
% Polyphase matrix is not applicable
PA=[];
otherwise
error('Specify valid method')
end
switch pph
case 3
lf0ea=length(F0even)+length(a);
lf1ed=length(F1even)+length(d);
lf0oa=length(F0odd)+length(a);
lf1od=length(F1odd)+length(d);
% This is tricky business, but it works
% -> for even- and odd length signals
% -> for even- and odd length filters
y0=repmat(0,floor(lxout/2),1);
y1=repmat(0,ceil(lxout/2),1);
xout=repmat(NaN,lxout,1);
y0(1:lf0ea-1)=conv(F0even(:),a(:));
y0(1:lf1ed-1)=y0(1:lf1ed-1)+conv(F1even(:),d(:));
y1(1:lf0oa-1)=conv(F0odd(:),a(:));
y1(1:lf1od-1)=y1(1:lf1od-1)+conv(F1odd(:),d(:));
% Even output
xout(even(xout),1)=y0;
% Odd output
xout(~even(xout),1)=y1;
case 2
% Synthesis
% Size of F0 is (lx+lh0-1+lf0-1) by (lx+lh0-1)
% Size of F1 is (lx+lf0-1+lh0-1) by (lx+lf0-1)
% and element (1,1) is h0(1)
% Filter Toeplitz matrices
F0=convmtx(f0(:),lx+lh0-1);
F1=convmtx(f1(:),lx+lf0-1);
% Construct the polyphase SYNTHESIS matrix
% Polyphase form
% Upsample: delete columns: keep these
F0=F0(:,d2hx);
F1=F1(:,d2fx);
% This already reduces the number of multiplications by two
% Now split in odd and even
xout=repmat(NaN,lxout,1);
F0even=F0(even(xout),:);
F0odd=F0(~even(xout),:);
F1even=F1(even(xout),:);
F1odd=F1(~even(xout),:);
PS=[F0even F1even ;
F0odd F1odd];
% Could make this TYPE II (p. 132)
% The output of this is mixed and it needs to be recombined!
% Check out the operation PS*PA*X - you've switched evens
% and odds and added a delay here and there so the
% reconstruction automatically makes up for it
xoutm=PS*ad;
% Output of the original evens of x
xout(even(xout),1)=xoutm(1:size(F0even,1));
% Output of the original odds of x
xout(~even(xout),1)=xoutm(size(F0even,1)+1:end);
case 1
% Synthesis
% Size of F0 is (lx+lh0-1+lf0-1) by (lx+lh0-1)
% Size of F1 is (lx+lf0-1+lh0-1) by (lx+lf0-1)
% and element (1,1) is h0(1)
% Filter Toeplitz matrices
F0=convmtx(f0(:),lx+lh0-1);
F1=convmtx(f1(:),lx+lf0-1);
% Upsample
blk0=~(1:lx+lh0-1);
blk1=~(1:lx+lf0-1);
blk0(d2hx)=a;
blk1(d2fx)=d;
w1=F0*blk0(:);
w2=F1*blk1(:);
xout=w1+w2;
% Polyphase is not applicable
PS=[];
otherwise
error('Specify valid method')
end
% Complete filter bank; length is shifted by the power of
% the only odd power of z^-1 that has a nonzero coefficient,
% i.e. by (length(P0)-1)/2
% Can also do wkeep(xout,size(x)).
xout=xout(l+1:end-l);
err=sqrt(mean((xout(:)-x(:)).^2));
if err>1e-10
error(sprintf('Reconstruction failed with %.2f',err))
else
disp(sprintf('FBANK Reconstruction error %8.4f',err))
end
disp(sprintf('FBANK (Analysis & Synthesis) took %8.4f s',etime(clock,t0)))
% Given the condition that we will choose 'conjugate mirror filters'
% F_0(z)=+H_1(-z)
% F_1(z)=-H_0(-z)
% to automatically satisfy the 'no-aliasing' requirement (4.2),
% the 'no-distortion' requirement (4.3) leads to finding
% P_0(z)=F_0(z)H_0(z)=z^-l where l is odd and even powers of z
% can be multiplied by any coefficient. We choose a lowpass
% P_0(z)=(1+z^-1)^2pQ(z) where Q(p) is a polynomial of degree 2p-2.
% Next, we can factor P_0 into different choices of H_0 and F_0,
% for various time lags l, and retrieve F_1 and H_1 from there.
% We take symmetric product filters, so l leads to 2l roots.
% The choice for l=1 is the Haar bank, l=3 leads to a 6th degree
% polynomial which can be split into two third degree polynomials,
% hence two 4-tap filters, called Daubechies-4. These are causal,
% but not linear phase.