Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Computing a DCT using an FFT (Read 20362 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Computing a DCT using an FFT

I am trying to make sense of the output of the FFT when used to calculate a type-2 DCT (DCT-II). I managed to scrape together enough documentation that infomed me that, if I wanted to perform a DCT-II using an FFT algorithm, then the data should be arranged to be real even symmetric, with every even indexed data element set to zero.

No problem there. In addition, the outputs are correct, but half the expected value (I can see the output needs scaling by 2 to compensate for the fact that every second sample of the input is zero).

However, I am tyring to figure out how the output comes to be what it is?
In particular, where the zero comes from (term 4 & term 12) [counting from 0].


Input

Real :  0  x0  0  x1  0  x2  0  x3  0  x3  0  x2  0  x1  0  x0
Imag :  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0

Output

Real :  ya  yb  yc  yd  0 -yd -yc -yb -ya -yb -yc -yd  0  yd  yc  yb
Imag :  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0

[grumble: looks good in fixed width font, but spaces are not turning out here]

In my real application, I am using a 16 point DCT, which equates to a 64 point FFT,  but the above is representative.

Any hints?



PS: I know there are better ways of doing DCTs, but in my implementation, I only have access to an FFT routine.


Cheers,Owen.

Computing a DCT using an FFT

Reply #1
Use code or codebox to get it looking right.

Do you mean you cannot do a rotation or similar before the FFT? This method looks very wasteful, there are much better ways of doing it via an FFT but they need a some pre- and postprocessing.

Computing a DCT using an FFT

Reply #2
Quote
Use code or codebox to get it looking right.


Thanks Garf,

I am not so worried about appearances, more wanting to understand why it is so. I am not that familiar with the mathematics, only how to apply the end results. What I expected, and I should have put in my original post, is...

I expected...
Real : ya yb yc yd -yd -yc -yb -ya -ya -yb -yc -yd yd yc yb ya
Imag : (all zero)

That is, no zero terms in the real output.

Quote
Do you mean you cannot do a rotation or similar before the FFT? This method looks very wasteful, there are much better ways of doing it via an FFT but they need a some pre- and postprocessing.
[a href="index.php?act=findpost&pid=348904"][{POST_SNAPBACK}][/a]


Yikes!! Not familiar with 'rotations' (but I have come across the term a lot). My application is targetted to hardware (so I can't use FP), and I am doing integer maths. I am trying to get away with 16 bits of precision (very convenient for the hardware).

I agree, the method is very wasteful, and I intend to use the radix-4 method of pre-processing to reduce the FFT size to n (rather than 4n). At the moment, I am trying out different things - but the output did give me a surprise. Non-mathematical documentation on doing a DCT using an FFT is sparse. I only found info on how the input is arranged [first post], but nothing on how the output is arranged.

I did try a 16-bit implementation of BinDCT, (internally, the precision extended to 22 bits) but the in-accuracies killed my application. I guess accuracy was the tradeoff for blistering speed!

I suppose my essential question is : is the output correct, or have I stuffed up? As mentioned, documentation was sparse, so I could be loading the FFT indicies wrong?

Thanks for the reply Garf.
Owen.

Computing a DCT using an FFT

Reply #3
I agree that there is often not enough documentation on how to do a simple DCT using an FFT.  Though it is not the best, computational-wise, it is however interesting to see the relationship between the two, esp. why the DCT is used more often than the FFT in compression (the even-symmetry creates smoothness that results in less high frequency energy).

Anyway, I was interested in this problem a while ago as well and found this link to be the best I've found.

Not sure if this is any help or not, but here is some MATLAB code I wrote during that time, which calculates the DCT using FFT.

Code: [Select]
clear all;
close all;

x=[1 2 3 4];
N=length(x);
d = dct(x);

x2=[4 3 2 1 1 2 3 4];
f=fft(x2);

n = 0:2*N-1;

shift=exp(-j*2*pi*(N-0.5)*n/(2*N));

f2 = real(f./shift);

d2 = f2(1:N)/sqrt(2*N);

d2(1)=d2(1)/sqrt(2);

disp(d);
disp(d2);


Notice how I don't do any zero-setting (no idea why you need to anyway).  I just take my original signal, mirror it and then take the fft.  Now the thing to note is that the Fourier transform of an even symmetry signal is real but we can't make true even symmetry signals on a computer, so we have to compensate for the phase shift caused by the time-shift of 3.5 samples, hence the division by the FFT of the variable 'shift'.  Then I scale the coefficients appropriately (to make them orthonormal), and voila.

Hope that helps in a little way 

EDIT:  fixed a spelling mistake

Computing a DCT using an FFT

Reply #4
Hi QuantumKnot,

Thanks for your reply. Indeed I had read the same page on using an FFT to do a DCT, and understood the reasoning behind having to mirror the co-efficents before placing them into a FFT.

I did a bit of experimenting, and created a "brute force" DCT using floating point maths on a PC. I had some code for a DCT-II, and re-wrote it to perform a DCT-I. The results were quite different.

I also examined the documentation on FFTW, in particular about the different types of DCT. They refer to them as 00, 01, 10, and 11. The 0 stands for no shift, the 1 stands for a half-sample shift. The first figure refers to the input, the last to the output. The FFTW routines REDFT00 = DCT-I, REDFT10 = DCT-II, REDFT01 = DCT-III, REDFT11 = DCT-IV.

Apparently, filling an FFT without padding but with mirroring (as you have done) is equivilant to a DCT-I.

Filling with padding and mirroring (as described in my first post) is supposed to be equivilant to a DCT-II (which is what I need). However, this takes the FFT to 4n. Further documentation (wikipedia in fact) hints at this "For example, a type-II DCT is equivalent to a DFT of size 4n with real-even symmetry whose even-indexed elements are zero. One of the most common methods for computing this via an FFT (e.g. the method used in FFTPACK and FFTW)..."
It goes on a bit further to say computations can be shorthened to O(nlogn) + O(n) if a pre-processing step of radix-4 is used (O(n) steps) and the FFT is then reduced to size n. A decent saving.

As I said before, my application is power impoverished hardware, where all I have is a 16-bit integer FFT routine (which works surprisingly well I must say). Additionally BinDCT may be very fast, but it is also considerably less accurate.

Again, thanks for your reply. A cursory look at your sample code has been enlightening, and I am about to print it and examine it in detail.

Cheers and thanks,
Owen.

Computing a DCT using an FFT

Reply #5
Quote
Apparently, filling an FFT without padding but with mirroring (as you have done) is equivilant to a DCT-I.
[a href="index.php?act=findpost&pid=350002"][{POST_SNAPBACK}][/a]


That is weird since I've also compared the results of the above MATLAB code with FFTW's REDFT10 r2r mode and got the same results (with the appropriate scaling).  If you check FFTW's user manual (p. 11), it says that a DCT-II is symmetric and even around j = -0.5, which is exactly how I'm mirroring it above.  That is, a sequence of abcd is mirrored as dcba|abcd, where if we align the right 'a' as 0, then the axis of symmetry (represented as a '|') is at -0.5.  The DCT-I is symmetric around j = 0, so the sequence abcd is mirrored as bcdabcd, where the axis of symmetry is on the 'a'.

Computing a DCT using an FFT

Reply #6
Quote
That is weird since I've also compared the results of the above MATLAB code with FFTW's REDFT10 r2r mode and got the same results (with the appropriate scaling).  If you check FFTW's user manual (p. 11), it says that a DCT-II is symmetric and even around j = -0.5, which is exactly how I'm mirroring it above.  That is, a sequence of abcd is mirrored as dcba|abcd, where if we align the right 'a' as 0, then the axis of symmetry (represented as a '|') is at -0.5.  The DCT-I is symmetric around j = 0, so the sequence abcd is mirrored as bcdabcd, where the axis of symmetry is on the 'a'.
[a href="index.php?act=findpost&pid=350009"][{POST_SNAPBACK}][/a]


That may explain where I am going astray. I was unsure how to perform the mirroring, so I just filled the FFT entries (excluding the zero padding, another issue) in the order abcddcba. An assumption I made (quite reasonable I think) is that since a DFT is periodic, this would affect either [1] the order of the output (when reading the same indexes), or [2] which outputs (indexes) of the FFT are relevant.

I did get my 4n size integer FFT working without any grief, and the results are (within 0.05% error) consistant with the original floating-point DCT. I have yet to code the radix-4 step, but from my reading so far, that seems fairly straight forward (just I wait!!).

It could well be that I can fill the entries properly (thanks very much for the explanation you have given above!!) and do a 2n FFT. I will give that a try...

Thanks,
Owen.

edit: OK, an update. I tried inserting values into the FFT (size 2n) as suggested. I was unhappy to find out that the outputs were nowhere near what was expected. Even worse, and more telling, the imaginary components of the output were not zero (as they should be if we shove real values in and want a DCT). Do you know if the "2n" method requires magnitude calculation of the real & imag outputs? (you don't mention anything about it in your post... I would imagine though that this would [should?] not be the case?).

Delving into the FFTW manual........

Cheers,
Owen.


Computing a DCT using an FFT

Reply #8
Quote
http://www.ee.columbia.edu/~marios/courses...%20(Thesis).pdf

Here is another paper for reference describing faster way's to compute a DCT using an FFT.
[a href="index.php?act=findpost&pid=356721"][{POST_SNAPBACK}][/a]


The paper is about the MDCT, not the DCT. And the MDCT is based on a type 4 DCT, while he needs a type 2 DCT.

Computing a DCT using an FFT

Reply #9
Quote
edit: OK, an update. I tried inserting values into the FFT (size 2n) as suggested. I was unhappy to find out that the outputs were nowhere near what was expected. Even worse, and more telling, the imaginary components of the output were not zero (as they should be if we shove real values in and want a DCT). Do you know if the "2n" method requires magnitude calculation of the real & imag outputs? (you don't mention anything about it in your post... I would imagine though that this would [should?] not be the case?).

Delving into the FFTW manual........

Cheers,
Owen.
[a href="index.php?act=findpost&pid=350255"][{POST_SNAPBACK}][/a]


Looks like I forgot to make a reply to this 

No, there is no need to do a magnitude calculation on the output.  If you mirror the sequence (size 2n) and do an FFT, you will NOT get pure real values, since there is a phase shift.  You need to correct the phase by multiplying the output with exp(-j*2*pi*(N-0.5)*n/(2*N)), and you will find that the output is now real (no imaginary component).

Computing a DCT using an FFT

Reply #10
% use FFT to commute DCT and then invert DCT using inverse FFT

x=1:6;

disp(x);

N=length(x);
d = dct  (x);

x2=[fliplr(x) x];
f=fft(x2);

n = 0:2*N-1;

shift=exp(-j*2*pi*(N-0.5)*n/(2*N));

f2 = real(f./shift);

d2 = f2(1:N)/sqrt(2*N);
d2(1)=d2(1)/sqrt(2);

disp(d)
disp(d2)

% now invert

d2(1)=d2(1)*sqrt(2);
d2=d2*sqrt(2*N);

f=[d2 0 -fliplr(d2(2:N))].*shift;
x2=ifft(f);

x=real(fliplr(x2(1:N)));
disp(x);

Computing a DCT using an FFT

Reply #11


I think the solution which is mentioned in below address in more optimized than your solution
because it uses N-point FFT to calculate N-point DCT , but you use 2*N point FFT:

http://fourier.eng.hmc.edu/e161/lectures/dct/node2.html

I have checked it in MATLAB it was okey except it has forgotten to use sqrt(2) and sqrt(2*N) coeficients as you use them.




clc;
clear all;
close all;

x = [ 1, 2, 3, 4, 3, 2, 1, 5, 1, 2, 3, 4, 3, 2, 1, 5, 1, 2, 3, 4, 3, 2, 1, 5, 1, 2, 3, 4, 3, 2, 1, 5 ];

matlab_idct = idct(x);

N=length(x);
n = 0:N-1;

shift = exp( (n*pi*j) / (2*N) );
shift = shift * sqrt(2*N);
shift(1) = shift(1) / sqrt(2);


y = shift .* x ;

y = real(ifft(y));

for i=0:((N/2)-1)
    my_idct(2*i + 1) = y(i + 1) ;
    my_idct(2*i + 2) = y(N - i) ;
end