ORTHOGONAL TRANSFORMS FOR DIGITAL SIGNAL PROCESSING

K. R. Rao N. Ahmed
Dept. of Electrical Engineering Dept. of Electrical Engineering
University of Texas at Arlington Kansas State University
ArI ington, Texas Manhattan, Kansas

Abstract

A tutorial—reviewpaper on discrete orthogonal transforms and their applications in digital signal and image
(both monochrome and color) processing is presented. Various transforms such as discrete Fourier, discrete cosine,
Walsh—Hadarnard, slant, Haar, discrete linear basis, Hadamard—Haar, rapid, lower triangular, generalized Haar,
slant Haar and Karhunen—Loeveare defined and developed. Pertinent properties of these transforms such as power
spectra, cyclic and dyadic convolution and correlation are outlined. Efficient algorithms for fast implementa-
tion of these transforms based on matrix partitioning or matrix factoring are presented. The application of these
transforms in speech and image processing, spectral analysis, digital filtering (linear, nonlinear, optimal and
suboptimal), nonlinear systems analysis, spectrography, digital holography, industrial testing, spectrometric
imaging, feature selection, and patter recognition is presented. The utility and effectiveness of these trans-
forms are evaluated in terms of some standard performance criteria such as computational complexity, variance
distribution, mean—square error, correlatedrms error, rate distortion, data compression, classification error,
and digital hardware realization.

Introduction for image data detection and extraction

and for image classification, The high energy compact-
In recent years there has been a growing interest re- ion property of the transform data has been taken ad-
garding the application of discrete orthogonal trans- vantage of in reducing the band width requirements
forms in the areas of digital signal and image process- (redundancy reduction), improving tolerance to channel
ing. This is primarily due to the impact of high speed errors, and achieving sample and/or bit reduction.
digital computers and the rapid advances in digital
technology and consequent development of special pur- Image processing by computer techniques in many cases
pose digital processors. Emergence of minicomputers, requires utilization of discrete orthogonal transforms.
microprocessors,and mini/micro systems has acted as a Fourier (DFT), slant (ST), Walsh—Hadamard (WHT), Hear
catalyst to the general field of data processing. Fast (HT), discrete—cosine (DCT), generalized Hear (GHT),
algorithms based on matrix factoring or matrix parti- fermat, slant—Haar (SHT), discrete linear basis (DLB),
tioning or other techniques resulting in reduced compu- Hadanard—Haar (HHT), rapid (RT), lower triangular, and
tational and memory requirements have further acceler- Karhunen—Love transforms (KLT) have already been
ated the uti lity and application of these transforms. utilized in some of the applications cited earlier.
Added advantage of these algorithms is the reduced Various performance criteria have been developed to
round—off error. This has resulted in a trend towards compare the uti I ity and effectiveness of these trans-
developing and standardizing the notation and termino- forms. The optimal transform is the KLT, since it de—
logy related to orthogonal functions and digital pro— correlates the transform coefficients, packs the most
cessi ng, energy in few coefficients, minimizes the mean—square
error (mse) between the reconstructed and original
Research efforts end related applications of orthogonal images and also minimizes the total entropy compared to
transforms include image processing, feature selection any other transform. But implementation of KLT in-
in pattern recognition, character recognition, signa- volves determinationof eigenvalues and corresponding
ture identificationand verification, characterization eigenvectors of the covariance matrix and there is no
of binary sequences, analysis and design of communica- general algorithm that enables its fast computational-
tion systems, digital filtering, spectral analysis, though some simplified procedures have been suggested
data compression, signal processing, convolution and for its implementationand some fast algorithms have
correlation processes, generalized Wiener filtering, been developed for certain class of signals. All the
spectrometric imaging, systems analysis, signal detect- other transforms possess fast algorithms for efficient
ion and identification, statistical analysis, spectro— computation of the transform operations, with some of
scopy and other areas. As such, orthogonal transforms them comparing fairly well in performance with that oF
have been used to process various types of data includ- the KLT,
ing speech, seismic, sonar, biological, biomedical,
forensic sciences, astronomy, satel I ite television pic- The objectives of this tutorial—reviewpaper are to de-
tures, aerial reconnaissance,weather photos, electron fine and develop the discrete orthogonal transforms and
nicrographs, range—Doppler planes, thermograms, X—rays, their properties, to develop the fast algorithms, to
two dimensional pictures of the human body, etc. The illustrate their applications and to compare their uti-
scope of interdisciplinarywork in this area of study lity and effectiveness in information processing based
is thus apparent. The importance and rapid appl ication on the standard performance criteria. It is hoped that
of digital techniques and digital processing can be ob- such a review of the state—of—the—artwould generate
served from the special issues devoted exclusively to additional interest and enthusiasm among all the engi—
these disciplines. neeringandscientific community and open further ave-
nues for research and appl icat ions.
Image processing involves spatial fi Itering, image cod-
Classification of Discrete
ing, image restoration and enhancement, image data ex-
traction and detection, color imagery, image diagnosis, Orthogonal Transforms
Wiener filtering, feature selection, pattern recogni-
The notion of signal representationcan be used as a
tion, digital holography, Kalman filtering, industrial
vehicle to categorize discrete orthogonal transforms.
testing, and others. Transform image (both monochrome
and color) coding has been utilized for image enhance- Signal representation implies the ability of a trans-
ment and restoration, form to efficiently represent a given set of data vect—

136

Authorized licensed use limited to: Horia Hulubei Natl Inst of Physics & Nuclear Eng. Downloaded on January 06,2023 at 15:44:29 UTC from IEEE Xplore. Restrictions apply.
ors in the form of an expansion, which consists of a called vecotr filtering as the elements of F(n)} are
finite number of terms. In order to compare the sig- not multiplied by scalars to yield {Z(n)}. When the
nal representationcapabilities of various transforms, transform is KLT the filter matrix is diagonal. How-
a criterion is necessary. One such often used criter- ever generation and implementationof KLT is a for-
ion is the rnse. The mae is minimum for the KLT which midable task. For other transforms the filter complex-
represents the eigenvectors of the covariance matrix ity increases, For suboptimal filters, [A(n)J can be
of a sampled data sequence0 The corresponding covar— constrained as follows:
iance matrix in the KLT domain is diagonal and its a) Constrain [A(n)J to be a diagonal matrix. This
nonzero elements are the eigenvalues Xi (il,2,...,N). iscalledscalarfiltering.
These represent the variances of the transform compo- b) Determine the filter design when the filter
nents. Data compression can be achieved by transmit- matrix cantains two elements per row and con-
ting only those components which have the largest var- tinue with additional matrix elements until the
iances. Thi.s can result in negligible mae, savings in desired mse is achieved. This approach, however,
channel capacity and in bit rate reduction. This pro- rapidly becomesquite complex.
cess of transform component selection is known as the To evaluate the effectiveness of various transforms in
variance criterion. suboptimal diagonal filtering the mse for a I order
Markov process when p=O.9 and (signal—to—noiseratio)
All the transforms other than the KLT are suboptimal as K01 is listed in Table 1.
they do not yield diagonal covariance matrices (The se-
quence is not completely decorrelated.) A reasonable Two—dimensional Wiener Filtering
way of classifying these transforms is to divide them
into two major categories, namely, U) optimal and (ii) Pratt also extended Wiener filtering to the two—dimen-
suboptimal. The latter can be divided into two more sional case. (Fig. 3) If the 2d—transformation is sep—
categories, Type 1 and Type 2. Type 1 consists of a arable,the transformation, its inverse and the filter-
class of transforms whose basis vector elements all lie ing can be carried out sequentially on the rows and
on the unit circle. Al I others will be considered as columns of the corresponding inputs. In matrix form
T1 TUZ(i,j) A
belonging to Type 2. Finally, Type 2 can be further
divided into two more categories, namely, Type 2—sinu-
soidal and Type 2—nonsinusoidal depending on whether
T
this can be expressed as EX(i,j)
Ai T where A
and Ai are the Wiener filter matrices
for the columns and rows of £F(u,v)]. This means the
the transform basis vector elements are sampled sinu- 2d—Wiener filtering can be implemented by sequential
soidal or nonsinusoidal functions respectively. This id—filtering on the rows and columns of EF(u,v)].
classification scheme is described in Fig. 1. Pratt has appi ied this filtering to noisy images using
WHT and has shown that there is negligible difference
Properties in the visual effects between the optimal and subopti—
rnal filtering. He has also applied sequency enhance-
As the transforms are orthogonal, fast algorithms are ment to images in the Walsh domain by using a linear
applicable to both forward and inverse transforms. filter function which compensates for the high sequen—
Also they can be applied to any number of dimensions cy attenuation.
and the order in which the transforms are applied to
multidirnensions can be interchanged. Also in most Log exp Filtering
cases multidimensional transform can be implemented
using id—transform in sequence. Only the DFT and RT For multiplicative noises a logarithmic preprocessor is
are invariant to the cyclic shift of the sequence0 first applied to the noisy data to separate the signal
Some of the other transforms, such as WHT, are invari- and noise components0 Linear filtering in the trans-
ant to the dyadic shift of the sequence. Shift invar- form domain can then be applied to minimize the noise
iant power spectra (both cyclic and dyadic) and convo- and interferenceeffects. This process is described in
lution and correlation properties (both cyclic and dy— block form in Fig. 4, Using WHT, Pratt has successful-
adic) for these transforms have also been developed, ly applied this technique in filtering a checker board
Transform techniques can be applied to convolution and type interferencepattern from an image. It may be
correlation processes and to digital filtering. worthwhile to investigatethe effectof scalar and vec-
tor filters with various transforms on image enhance—
App I ications ment.

Efficient algorithms for the rapid implementationof Nonlinear Filtering

the discrete orthogonal transforms resulting in re-
duced memory and round off error have acted as a cata- Pratt has suggested a number of nonlinear filtering
lyst in their application to diverse fields ranging operations in the transform domain to improve the image
from signal processing to pattern recognition. Some of quality in terms of edge enhancement, sharpening the
these applications have been cited in the introduction. image detail etc. One method is to raise the transform
The object here to summarize some of these application component to a power 1 + c where s is a smal I number.
areas. Another process is a logarithmic operation (Fig. 5).
Filter design based on various performance criteria re-
Generalized Wiener Filtering quires further investigation.

Pratt has extended the classical Wiener filtering tech- Seguency Domain Desiqpf Frequency Filters
nique to transform processing (Fig. 2). The filtering
operation can be implemented by various orthogonal Kahveci and Hall have developed equivalent filtering
transforms and is also applicable to two dimensional techniques for implementing a filter in one transform
data such as images. The object here is t2 select the domain given the filter in another transform, such as
x 2n> filter matrix [A(nYl such that X(n)} the Fourier, domain (Fig. 6). They have illustrated this
estimate of fX(n) is the best in the mse sense when method with a Walsh filter, which is computationally
the signal {X(n)} has been corrupted by an additive simpler but still equivalent to a Fourier filter which
noise {W(n)}. The filter matrix depends on the stat- is easier to design. This technique has also been ex-
istics of the signal and the noise and in general it is tended to the two—dimensionalcase. Possible decompo-
not a diagonal matrix. This type offiltering is sition of the two—dimensional filter into two (row and

137

Authorized licensed use limited to: Horia Hulubei Natl Inst of Physics & Nuclear Eng. Downloaded on January 06,2023 at 15:44:29 UTC from IEEE Xplore. Restrictions apply.
 column) one—dimensional filters and equivalent filter IAVZS), geometrical zone etc. Additional reduction in
design in Haar, slant and other transform domains are bit rates can be achieved by adopting variable bit as-
worth looking into. signment i.e., bit allocation based on variance, magni-
tude, log of the variance, log of the threshold etc.
Feature Selection The object in all of these cases.is to reconstruct the
image with negligible degradation and maximum savings
As the components tend to be less correlated in the in channel capacity. Image enhancement can be effected
transform domain in comparison to the original signal by noise filtering and by introducing weights on the
domain, with complete decorrelation for the KLT, it is transform components. Hybrid schemes (transform—DPCM)
possible to simplify the problem of pattern recognition also have been adopted. These result in a compromise
with reduced number of features. One basis for select- of the two techniques. Real time image processing for
ing the features in the transform domain is the variance video transmission using DFT, WHT and HT has also been
criterion. By choosing those components that have the carried out. WHT, HHT, and ST have been mpplied to
largest variances, pattern recognition based on reduced image data coding. The various image processes are
number of features with negl igible misclassification described in block form in Figs. 10 through
can be achieved. Such a technique has been developed
by Andrews for recognizing numeral characters. (Fig. 7) Conclusions

Ahead et al have applied this technique to classifying Fast algorithms together with the developments in digi-
canine ECG data as normal or abnormal and have shown tal computers and in other digital hardware have
that based on the variance criterion a 3:1 data com- spurred the rapid application of discrete transforms in
pression can be achieved without loss of the features the general areas of signal and image processing.
necessary for the physician to interpret the ECG. Hence Special purpose digital processors tailored for specific
the variance distribution can be a criterion for the projects based on selected transforms have been built
effectiveness of the transform. For purposes of com- while a number of transforms have been developed, It
parison, such a distribution is shown for a I order Mar— appears only a few of them such as DFT, WHT, HT, ST,
key process with p = 0.9, K0 1 and N = 16 for various OCT and RT have significant advantages in practical
transforms in Table 2. applications. It is anticipated that the utility and
effectiveness of these transforms in a number of diverse
Wang and Shiau have also used the variance criterion for fields will be further accelerated.
selecting features in recognizing single—font printed
Chinese characters. They have achieved dimensionality
reduction in feature space using the OFT, WHT and RT.
Their computer simulation has showed that the RT is the
I Distrete Orthogonal Transformsj
.1.

most attractive one in unusual situations such as shift-

ed and/or rotated patterns and also because of its fast Optimal Suboptimal Transforms I

implementation. The transform technique has led to ex- Karhunen—Loeve

Transforms]
tracting fewer but relevant features resulting in sim- 'V
plified classifier design Further research is needed LTYPU Type 2
in developing an automatic machine recognition and
translation system.
I
Discrete Fourier Type 2—
.1,

Type 2-Sinusoidal
Rate Distortion Waish-Hadaisard Nonsinusoidal Discrete Cosine
Generalized Haar
Another criterion for evaluating the orthogonal trans- Continuous Slant
forms is the rate distortion. The rate distortion Generalized Discrete
function R(D) yields the minimum information rate in Linear Basis
bits per transform component needed for coding such that Hadamard—Haar
the average distortion is lass than or equal to a cho- Slant—Haar
sen value D for any specified source probability distri- Fermat
bution. The rate distortion is a measure ot the decor— Generalized
relation of the transform components as the distortion Hear
can be spread uniformly in the transform domain thus
minimizing the rate required for transmitting the in- Fig. 1. A Classification of Discrete Orthogonal
formation. The rate versus distortion for a I order Transforms
Markov process for N=l6 and p=O.9 is shown in Fig. 8
for a number of transforns. Inspection of this figure
shows that in terms of the rate distortion KLT is the
best with the DOT and DFT being very close to it fol-
lowed by (HHT)1, (HHT)2, HT and WHT. The identity [F(n)J
transform is the least favorable as it maintains the
correlation in the signal. tWforff n)} [A ri)

Image Processing
—
{X(n)}
The techniques of bandwidth reduction, data compression J Inverse
and noise filtering applicable to a sequence can be ex- 1 Transform,T -'1

tended to digitized images (both color and monochrome).

Transform image processing is destribed in block form
in Fig. 9. Sample and/or bit reduction is achieved by Fig. 2. Generalized Wiener Filtering
transmitting only a fraction of the transform components
that have the most energy. Selection of these components
can be based on a variety of criteria such as magnitude
(threshold), variance (maximum variance zonal sampling—

138

Authorized licensed use limited to: Horia Hulubei Natl Inst of Physics & Nuclear Eng. Downloaded on January 06,2023 at 15:44:29 UTC from IEEE Xplore. Restrictions apply.
 N NT WHT CHT DFT DCT ST KLT (SdT)1 (HUT)1 (HUT)2
2 0.3730 0.3730 0.3730 0.3730 0.3730 0.373000 0.3730 0.3730 0.3730 0.3730
4 0.2942 0.2942 0.2942 0.2964 0.2920 0.292092 0.2915 0.2942 0.2942 0.2942
8 0.2650 0.2649 0.2650 0.2706 0.2456 0.256206 0.2533 0.2595 0.2649 0.2649
16 0.2589 0.2582 0.2589 0.2592 0.2374 0.243423 0.2356 0.2534 0.2584 0.2582
32 0.2582 0.2582 0.2582 0.2441 0.2282 0.240234 0.2268 0.2576 0.2568
64 0.2581 0.2559 0.2580 0.2571
Table 1. Mean square error for first order Markov process when P = 0.9 and = 1 for
scalar Wiener filtering

Transform
j HT CUT WUT DCT DFT ST (SHT)1 (HUT)1 (HHT)2
1 8346 9.8346 9.8346 9.8346 9.8346 9.8346 9.8346 9.8346 9.8346
2 2,5364 2.5364 2.5360 2.9328 1.8342 2.8536 2.7765 2.5364 2.5364
3 0.8638 0.8635 1.0200 1.2108 1.8342 1.1963 1.0208 1.0209 1.0209
4 0.8638 0.8635 0.7800 0.5814 0.5189 0.4610 0.4670 0.7061 0.7061
5 0.2755 0.2755 0.7060 0.3482 0.5189 0.3468 0.3092 0.2946 0.3066
6 0.2755 0.2755 0.3070 0.2314 0.2502 0.3424 0.3031 0.2946 0.3031
7 0.2755 0.2755 0.3030 0.1684 0.2502 0.1461 0.2837 0.2562 0.2864
8 0.2755 0.2755 0.2830 0.1294 0.1553 0.1460 0.2059 0.2562 0.2059
9 0.1000 0.1000 0.2060 0.1046 0.1553 0.1047 0.1042 0.1024 0.1038
.0 0.1000 0.1000 0.1050 0.0876 0.1126 0.1044 0.1042 0.1024 0.1038
.1 0.1000 0.1000 0.1050 0.0760 0.1126 0.1044
0.1034 0.1024 0.1034
.2 0.1000 0.1000 0.1040 0.0676 0.0913 0.1034 0.0631
0.1024 0.1034
.3 0.1000 0.1000 0.1040 0.0616 0.0913 0.1010 0.0976 0.1013
0.0631
14 0.1000 0.1000 0.1030 0.0574 0.0811 0.0631
0.1010 0.0976 0.1013
.5 0.1000 0.1000 0.1020 0.0548 0.0811 0.0913 0.0976 0.0913
0.0631
6 0.1000 0.1000 0.0980 0.0532
0.0780 0.0631 0.0913 0.0976 0.0913
Table 2. Variance distribution for a I order Markov process, p = 0. 9 and
= 1
K0
'rarisform Number of Arithmetic Operations Recoil re fl (x) _____ log
f (x)+log fn (x)I ld—
HT 2(N — 1) S
lTransforrn
RT N log2N
NHT N log2N I
Z(u)=Z s (u)+Zn (u)
ld-inverse ____ (x)
— 1) +N = (3N — transform
l
( exp I
(HUT)
1 4( Real
I

(HUT) 2 8 (— 1) + 2N— (4N— 8) Linear Filter

Fig. 4. log exp transform filtering
(HHT) 2
r+l N
(—i —
1) + rN = [(r + 2)N — 2r+l]
(SUT) (r + 2)N— 2r+l
f(x) ld— F(u) 'nonlinear filter
ST N log2N + (2N 4)
—

transform1
CUT

DFT
3M - 4
N log2N
g[l+KF(u)IF(u) I

Complex
DCT 2N log2 (N)
hi ld-inversel f(x)
KLT N2 ] '1 transform

Table Number of arithmetic operations (real

3. com- or Fig. 5. Nonlinear Transform Filtering.
K1 and K2
plex) required for fast implementationof various are scaling constants
discrete transforms. An arithmetic operation is corn—
posed of a multiplication and an addition discrete
(subtraction). Note that NT and WHT require additions
only. KLT has no fast implementaion but is included
for purposes of comparison. The arithmetic opera-
tions required for the KLT can be real or complex Inverse
f
depending on the covariance matrices. transform
]ransform I

a) Gt
A
f =
{Z(i,j) ]=[X(i,j) }+[W(i,j) 3 2d—
Transform
{F(u,v)

b)
Transfoi] IJG I Inverse
Pj transform F
I
1 fB

Fig. 6. Equivalent filtering. a)Filtering specified

in transform A domain. b)Equivalent filter-
filter
ing in transform B domain.
Fig. 3. Two-dimensionalGeneralized Wiener Filtering

139

Authorized licensed use limited to: Horia Hulubei Natl Inst of Physics & Nuclear Eng. Downloaded on January 06,2023 at 15:44:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 7. Numeral character recognition using variance
criterion for feature selection

Q2
—0
04 oe 8 t.0

Fit. 8 RakVenusdistortion forM 16 and p =0.9.

Figure 9. Description of various transform and/or DPCM processes adopted for image data processing

Re
ldtransform hannel Decodin
MVZ

Figure 10. DPCM—Transformprocessing using threshold sampling or MVZS

Reconstructed

Figure 11. Transform-DPCMprocessing using threshold sampling or NVZS

Original Reconstructed

Figure 12. ld—transform processing using MSJZS or threshold sampling and spectral extrapolation

Original
ld—transform Selector: ld-inverse Reconstructed
along rows MVZS or threshold transform
.

image j
image
Figure 13. ld—transform processing using maximum variance zonal sampling or threshold sappling

140

Authorized licensed use limited to: Horia Hulubei Natl Inst of Physics & Nuclear Eng. Downloaded on January 06,2023 at 15:44:29 UTC from IEEE Xplore. Restrictions apply.