Professional Documents
Culture Documents
SUMMARY
The discrete Hartley transform (DHT) is a real-valued transform closely related to the discrete Fourier
transform (DFT) of a real-valued spectrum which preserves some useful properties of DFT. In this paper,
an algorithm is proposed to realize one-dimensional DHT with sequence length N = 2(2m) , m2. The
proposed algorithm requires fewer multiplications than some of the algorithms in the literature. Copyright
q 2008 John Wiley & Sons, Ltd.
Received 22 July 2007; Revised 7 October 2008; Accepted 11 October 2008
KEY WORDS:
1. INTRODUCTION
Over the years, the discrete Hartley transform (DHT) [1] has been established as a potential tool for
signal processing and communication applications, e.g. computation of convolution and deconvolution [25], interpolation of real-valued signal [6], image compression [711], error control coding
[12], adaptive filtering [1316], multi-carrier modulation and many other applications [1720].
The DHT is known due to its real-valued symmetric transform kernel, which is identical to that
of its inverse. It is used as a real-valued alternative to the discrete Fourier transform (DFT) in
optical holography and phase measurement [21]. Recording of the optical intensity of the Hartley
transform amounts only to discarding of the sign, whereas recording the intensity of the Fourier
transform corresponds to discarding the phase, resulting in a much more serious loss of information
[22]. Since the introduction of the fast Hartley transform (FHT) by Bracewell in 1984 [1], an
intense research effort has been made by several investigators leading to a number of algorithms
[2334]. The direct calculation of the DHT is computationally intensive and requires N 2 multiplications and additions [1]. Such computational complexity is costly in terms of reduced signal
Correspondence
to: Priyanka Jain, Indira Gandhi Institute of Technology, I P University, Kashmere-Gate, Delhi110006, India.
E-mail: priyajain2000@rediffmail.com
Copyright q
410
processing speed, increased power consumption and higher expense. Overall, these methods are
focused either on the minimization of the total computational complexity [5, 23, 28, 29] or on the
minimization of the multiplicative complexity [24, 25, 27, 30].
The basic idea of the proposed algorithm implies a recursive decomposition technique. By using
the recursion for N = 2(2m) , m2 (this will ensure that N 16), the DHT is eventually decomposed
into a N /2 DHT and a number of N /4 lengths discrete Cosine transforms (DCT). Here, recursive
structures of folded DCTs are used [35]. We extend this approach for the case N 22m and show
that the suggested methodology results in the reduced multiplications as compared with those
reported in [36, 37].
2. FACTORIZATION OF DHT
A sequence x[n], n = 0, 1, . . . , N 1 has a DHT defined by [34]
H [k] =
N
1
x[n] cas
n=0
2kn
,
N
k = 0, 1, . . . , N 1
(1)
where cas cos +sin . Let us define two additional sequences H1 [k] and H2 [k] by
H1 [k] =
N /21
n=0
H2 [k] =
N /21
n=0
2kn
,
x[2n] cas
N /2
k = 0, 1, . . . , N /21
2(2n +1)k
,
x[2n +1]cas
N
(2)
k = 0, 1, . . . , N /21
(3)
Here H1 [k] and H1 [k] are the DHT of size N /2 applied to the even and odd samples of x[n],
respectively. H2 [k] can be written as
H2 [k] =
N /41
x[2n +1]cas
n=0
N /41
(2n +1)k
(N 2n 1)k
+
x[N 2n 1]cas
2(N /4)
2(N /4)
n=0
k = 0, 1, . . . , N /21
(4)
or
H2 [k] =
N /41
n=0
N /41
(2n +1)k
(2n +1)k
+
x[2n +1]cas
x[N 2n 1]cms
2(N /4)
2(N /4)
n=0
k = 0, 1, . . . , N /21
Copyright q
(5)
Int. J. Circ. Theor. Appl. 2010; 38:409417
DOI: 10.1002/cta
411
where cms cos sin . Combining the cosine and sine terms into two separate summations,
we obtain
(2n +1)k
(x[2n +1]+ x[N 2n 1]) cos
H2 [k] =
2(N /4)
n=0
N /41
(2n +1)k
,
(x[2n +1] x[N 2n 1]) sin
+
2(N /4)
n=0
N /41
k = 0, 1, . . . , N /21
G 1 [k]+ G 2 [k]
(6)
(7)
where the first series on RHS of (6) is defined as G 1 [k] and the second as G 2 [k]. We observe that
G 1 [k] is a DCT of length N /4, without the normalization factor 1 for the dc term.
2
By replacing k by (N /4k) in G 2 [k] above, we get
G 2 [N /4k] =
N /41
(1)n [x[2n +1] x[N 2n 1]] cos
n=0
(2n +1)k
2(N /4)
k = 0, 1, . . . , N /41
(8)
which is again a DCT of length N /4. Note that G 1 [k] and G 2 [N /4k] are N /4 point DCTs of
addition and subtraction of the folded sequence comprising odd samples of x[n], respectively. The
following properties of the sequences H1 , H2 , G 1 and G 2 can be used for further simplification:
H1 [k+N /2] = H1 [k],
H2 [k + N /2] = H2 [k],
G 2 [N /2k] = G 2 [k],
G 1 [N /4] = 0,
G 1 [N /2k] = G 1 [k]
G 2 [0] = 0
(9)
Using the above equations we can compute H [k] for 0kN 1 from H1 [k], 0kN /21, G 1 [k]
and G 2 [k], 0kN /41.
A general flow graph of the decomposition above is shown in Figure 1. For an N -point DHT,
we need one DHT of size N /2 and two DCTs of size N /4. DCTs can be calculated from recursive
filter structure [38]. It is well known that the recursive architecture is regular, modular, and free
of global interconnection.
N /41
n=0
(2n +1)k
2(N /4)
(10)
1
N
x[n]
cos
n=0
Copyright q
(2n +1)k
,
2N
k = 0, 1, 2, . . . , N 1
(11)
412
H[k]
x[n]
0
1
0
2
N/2
DHT
H1[k]
N-2
N/2-1
H2[k]
1
3
7
5
15
13
9
11
N/2
N/4
DCT
G1[k]
N/4
DCT
G2[k]
N-1
Figure 1. Factorization of the DHT into one DHT and two DCTs (flow diagram N = 16).
N
/21
w k [n] cos
n=0
(2n +1)k
,
2N
k = 0, 1, 2, . . . , N 1
(12)
where
k
x[N
n 1])
w k [n] = (x[n]+(1)
(13)
(14)
The derivation of the kernel for G 2 [k] is similar, and we briefly write the results here for completeness. Using (8), replacing N /4 by N , we obtain
1
N
(2n +1)k
, k = 0, 1, 2, . . . , N 1
x[n]
cos
(15a)
G 2 [N k] =
2(N )
n=0
where
x[n]
= (1)n (x[2n +1] x[N 2n 1])
After folding x[n]
at
N /2,
G 2 [N k] =
N
/21
n=0
Copyright q
(15b)
w k [n] cos
(2n +1)k
2N
(16a)
413
where
k
x[N
n 1])
w k [n] = (x[n]+(1)
(16b)
(17)
N
/21
w k
n=0
N
n
2
(2n +1)k
(k1)/2
(1)
,
sin
2N
k odd
(18)
i.e.
G 1 [k] = (1)((k1)/2) PN /21 [k],
where
Pi [k] =
k = 1, 3, . . . , N 1
1
k ,
w k [i n] sin n +
2
n=0
i
(19)
k odd
(20a)
and
k
N
We use the following trigonometric identities for further simplification:
k =
Pi [k] =
i
(21a)
(21b)
(22)
n=0
= 2 cos k
i1
k
sin w k [i]+
w k [i 1n] sin[(n +1/2)k ]
2
n=0
i2
k
3k
w k [i]sin w k [i 1]+
w k [i 2n] sin[(n +1/2)k ]
sin
2
2
n=0
= 2 cos k sin
Pi [k] = sin
Copyright q
(20b)
(23)
k
3k
k
w k [i]+ sin
w k [i]+2 cos k Pi1 [k]+ sin w k [i1]Pi2 [k] (24)
2
2
2
k
{w k [i]+ w k [i 1]}+2 cos k Pi1 [k] Pi2 [k],
2
k odd
(25)
414
k = 1, 3, . . . , N 1
(26a)
where
Ri [k] =
i
w k [i n] sin(n +1/2)k ,
k odd
(26b)
n=0
Ri [k] = sin
k
{w k [i]+ w k [i 1]}+2 cos k Ri1 [k] Ri2 [k],
2
k odd
(27)
k = 0, 2, . . . , N 2
(28)
where
1
k ,
w k [i n] cos n +
Q i [k] =
2
n=0
i
k even
(29)
k
{w k [i] w k [i 1]}+2 cos k Q i1 [k] Q i2 [k],
2
k even
(30)
k = 0, 2, . . . , N 2
(31)
where
Si [k] = cos
k
{w k [i] w k [i 1]}+2 cos k Si1 [k] Si2 [k],
2
k even
(32)
4. PERFORMANCE
Using (25), (30) and (27), (32), the DCTs: G 1 [k] and G 2 [k], respectively, can be realized with an IIR
filter structure, as shown in Figure 2. We then obtain H2 [k] = G 1 [k]+ G 2 [k] (due to (7)). H1 [k] is
obtained as N /2 point DHT of x[2n] samples. Thereafter, the complete DHT: H [k] = H1 [k]+ H2 [k]
is computed. A detailed calculation of the multiplications needed for the computation of the DHT
for sequence x[n] of length N shows that these are less than those proposed in [36, 37], as shown
in Table I. In addition, the additional advantage of the proposed methodology is the multiplications
needed in splitting the computation of DCT in four recursive structures, which can be evaluated
simultaneously with an overall reduced computational time.
Copyright q
415
x[n].
Proposed
Reference [36]
Reference [37]
6
18
58
202
14
34
74
258
40
104
256
5. CONCLUSION
A novel method for fast computation of DHT has been proposed. We use DHT of N /2 samples
combined with two DCTs of N /4 samples. DCT computation is further realized by folding N /4
samples, i.e. using N /8 samples. This compares favorably to the earlier available designs in the
literature [36, 37]. Mathematical derivation of the proposed computation has been given. Hardware
structure resulting from the suggested algorithm has also been given.
ACKNOWLEDGEMENTS
The authors are indeed obliged to the reviewers for their highly constructive suggestions, which improved
the quality of the paper.
REFERENCES
1. Bracewell RN. The fast Hartley transform. Proceedings of the IEEE 1984; 72:10101018.
2. Meher PK, Patra JC, Swamy MNS. High-throughput memory-based architecture for DHT using a new
convolutional formulation. IEEE Transactions on Circuits and SystemsII: Express Briefs 2007; 54(7):606610.
3. Bracewell RN. The Fourier Transform and its Applications (3rd edn). McGraw-Hill: New York, 2000.
Copyright q
416
4. Cheng L-Z, Tong L, Jiang ZR. Real transform algorithm for computing discrete circular deconvolution. Proceedings
of the 3rd IEEE International Conference on Signal Processing, vol. 1, Beijing, China, October 1996; 166169.
5. Duhamel P, Vetterli M. Improved Fourier and Hartley transform algorithms: application to cyclic convolution of
real data. IEEE Transactions 1987; ASSP 35(6):818824.
6. Maharana GS, Meher PK. Algorithm for efficient interpolation of real-valued signals using discrete Hartley
transform. Computers in Electrical Engineering 1997; 23(2):129134.
7. Tzou KH, Hsing TR. A study of discrete Hartley transform for image compression application. Proceedings of
SPIE International Society for Optical Engineering 1985; 534:108.
8. Paik CH, Fox MD. Fast Hartley transform for image processing. IEEE Transactions on Medical Imaging 1988;
7(6):149153.
9. Duleba I. Hartley transform in compression of medical ultrasonic image. Tenth International Conference on
Image Analysis and Processing (ICIAP99), Venice, Italy, 1999; 722.
10. Bracewell RN. Affine theorem for the Hartley transforms of an image. Proceedings of the IEEE 1994; 82(3):
388390.
11. Sunder RS, Eswaran C, Sriram N. A 3-D discrete Hartley transform coder for compression of magnetic resonance
images. IEEE International Conference on Electro Information Technology, Lincoln, U.S.A., 2225 May 2005; 6.
12. Wu J-L, Shiu J. Discrete Hartley transform in error control coding. IEEE Transactions on Signal Processing
1991; 39(10):23562359.
13. Meher PK, Panda G. Unconstrained Hartley-domain least mean square adaptive filter. IEEE Transactions on
Circuits Systems II, Analog Digital, Signal Processing 1993; 40(9):582585.
14. Muralishankar R, Sangwan A. Complex cepstrum of discrete Hartley and warped discrete Hartley filters. IEEE
12th Signal Processing Education Workshop, vol. 4, Teton National Park, WY, U.S.A., September 2006; 392397.
15. Muralishankar R, Shankar HN, Shaughnessy DO. A performance analysis of features from complex cepstra
of warped DST, DCT and DHT filters for phoneme recognition. Proceedings of the IEEE 15th International
Conference on Digital Signal Processing (DSP 2007), Cardiff, Wales, U.K., 591594.
16. Tseng C-C, Lee SL. Closed-form design of fractional delay FIR filter using discrete Hartley transform. IEEE
Region of 10th Conference 2007 (TENCON 2007), Taipei, 30th October2nd November 2007; 14.
17. Villasenor JD. Optical Hartley transform. IEEE Proceedings 1994; 82(3):391399.
18. Popovic M, Sevic. A new look at the comparison of the fast Hartley and Fourier transform. IEEE Transactions
on Signal Processing 1994; 42(8):21782182.
19. Wang D, Liu D, Liut F, Yue G. A novel DIT-based ultra-wide band system. Proceedings of the IEEE International
Symposium on Communications and Information Technology (ISCIT 2005), vol. 1, Beijing, China, 1214 October
2005; 672675.
20. Lin Z, Yin F, McCallum RW. A fast running Hartley transform algorithm and its application in adaptive signal
enhancement. IEEE International Conference on ASSP (ICASSP 2005), vol. 4, Philadelphia, PA, U.S.A., 1823rd
March 2005; iv/189iv/192.
21. Bracewell RN. Aspect of the Hartley transform. IEEE Proceedings 1994; 82(3):381387.
22. Millane RP. Analytic properties of the Hartley transform and their implication. IEEE Proceedings 1994; 82(3):
413428.
23. Sorensen HV, Jones DL, Burrus CS, Heideman MT. On computing the discrete Hartley transform. IEEE
Transactions on Acoustics, Speech, Signal Processing 1985; ASSP-33:12311238.
24. Bi G. Split-radix algorithm for the discrete Hartley transform. Electronics Letters 1994; 30:18331835.
25. Bouguezel S, Ahmad MO, Swamy MNS. A new radix 2/8 FHT algorithm for length q 2m DFTs. IEEE
Transactions on Circuits and Systems 2004; 51(9):17231732.
26. Bouguezel S, Ahmad MO, Swamy MNS. A split vector-radix algorithm for the 3-d discrete Hartley transform.
IEEE Transactions on Circuits and SystemsI 2006; 53(9):19661976.
27. Bi G, Chen YQ. Fast DHT algorithms for length N = q 2m . IEEE Transactions on Signal Processing 1999;
47:900903.
28. Jonckheere EA, Ma CW. Split-radix fast Hartley transform in one and two-dimensions. IEEE Transactions on
Signal Processing 1991; 39(2):499503.
29. Chan YH, Siu WC. New fast discrete Hartley transform algorithm. Electronics Letters 1991; 27:347349.
30. Hou HS. The fast Hartley transform algorithm. IEEE Transactions on Computer 1987; C-36:147156.
31. Bi G. New split-radix algorithm for the discrete Hartley transform. IEEE Transactions on Signal Processing
1997; 45(2):297302.
32. Kumar NV, Prabhu KMM. FHT algorithm for length N = q.2m . Electronics Letters 1999; 35(12):966968.
Copyright q
417
33. Bi G, Chen YQ, Zeng Y. Fast algorithms for generalized discrete Hartley transform of composite sequence lengths.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing 2000; 47(9):893901.
34. Jones KJ. Design and parallel computation of regularized fast Hartley transform. IEE ProceedingsVisual, Image
Signal Process 2006; 153(1):7078.
35. Chen CH, Liu BD, Wang JL, Yang JF. Efficient recursive structure for forward and inverse discrete Cosine
transform. IEEE Transactions on Signal Processing 2004; 52(9):26652669.
36. Kober V. Fast algorithms for computation of sliding discrete Hartley transforms. IEEE Transactions on Signal
Processing 2007; 55(6):29372944.
37. Shu H, Wang Y, Senhadji L, Luo L. Direct computation of type-II discrete Hartley transform. IEEE Processing
Letters 2007; 14(5):329332.
38. Jain P, Kumar B, Jain SB. Discrete sine transform and its inverse-realization through algorithms. International
Journal of Circuit Theory and Applications 2008; 36(4):441449.
Copyright q