Wiener discrete cosine transform-based

image filtering

Oleksiy Pogrebnyak
Vladimir V. Lukin
Journal of Electronic Imaging 21(4), 043020 (Oct–Dec 2012)

Wiener discrete cosine transform-based image filtering

Oleksiy Pogrebnyak
Instituto Politecnico Nacional
Centro de Investigacion en Computacion
Av. Miguel Othon Mendizabal s/n, Col. La Escalera, Del. Gustavo A. Madero
C.P. 07320 Mexico City, Mexico
E-mail: [email protected]

Vladimir V. Lukin
National Aerospace University
Department 504, 17 Chkalova Street
61070 Kharkov, Ukraine

of potential limit output (PLO) mean square error (MSE) for

Abstract. A classical problem of additive white (spatially uncorre-
lated) Gaussian noise suppression in grayscale images is considered. grayscale (one-component) images has been put forward by
The main attention is paid to discrete cosine transform (DCT)-based Chatterjee and Milanfar.7 This approach presumes that noise
denoising, in particular, to image processing in blocks of a limited size. is AWGN and a noise-free image is available. Later, this
The efficiency of DCT-based image filtering with hard thresholding is approach has been further advanced8 to allow predicting the
studied for different sizes of overlapped blocks. A multiscale approach PLO MSE without having a quite accurate corresponding
that aggregates the outputs of DCT filters having different overlapped
block sizes is proposed. Later, a two-stage denoising procedure that noise-free image.
presumes the use of the multiscale DCT-based filtering with hard The results presented in Refs. 8–10 demonstrate the fol-
thresholding at the first stage and a multiscale Wiener DCT-based lowing: for a given image, the PLO MSE decreases if noise
filtering at the second stage is proposed and tested. The efficiency variance reduces. For a given noise variance, the PLO MSE
of the proposed multiscale DCT-based filtering is compared to the can vary by several times depending upon an image. It can be
state-of-the-art block-matching and three-dimensional filter. Next,
the potentially reachable multiscale filtering efficiency in terms of out-
easily concluded from data presented in Ref. 7 that the PLO
put mean square error (MSE) is studied. The obtained results are of MSE is considerably, by up to 10 times, larger for more com-
the same order as those obtained by Chatterjee’s approach based on plex structure (highly textural) images. Within the approach
nonlocal patch processing. It is shown that the ideal Wiener DCT- in Ref. 7, the PLO MSE is practically reached by modern
based filter potential is usually higher when noise variance is high. most efficient filters for complex-structure images.
© 2012 SPIE and IS&T. [DOI: 10.1117/1.JEI.21.4.043020] The PLO MSE in Ref. 7 has been derived within a non-
local filtering approach. There are many techniques that
belong to this family nowadays. They are based on searching
1 Introduction for similar patches and their joint processing.11–14 Among
Noise is one of the main factors that degrades image qual- them, the block-matching three-dimensional (BM3D) filter14
ity.1,2 In spite of considerable efforts spent on noise intensity has been shown to be the most efficient for processing most
reduction in originally acquired images, noise still remains grayscale test images7 and component-wise denoising of
visible and disturbing for many practical applications. There color test images10 corrupted by AWGN.
are different types of noise that can be present in images Meanwhile, the approach in Ref. 7 might not be unique
such as additive white Gaussian noise (AWGN), spatially for determination of PLO MSE. From the linear filtering
correlated additive noise, signal-dependent and mixed noise, theory, the Wiener filter is known to be the optimal in the
speckle, etc.3–6 And there are various groups of methods sense of providing minimal output MSE under the condition
for image denoising. However, researchers continue their of a priori known spectra of stationary signal and noise.15
attempts to design new, more efficient techniques for both Wiener filtering being applied to processing an entire image
quite general and more specific applications. in spatial two-dimensional (2-D) Fourier domain is not as
One reason is that the image processing community and efficient as in the case of one-dimensional (1-D) stationary
customers are not satisfied by the already obtained results. signal filtering (stationarity is required for proper operation
Another reason is that until recently it has not been clear of the Wiener filter16), since images are nonstationary ran-
that there is room for further improvement of image filtering dom 2-D processes. Because of this, quasi-Wiener filtering
performance. Fortunately, a new approach to the estimation is often implemented in spatial domain locally. The widely
known local statistic Lee17 and Kuan18 filters are good exam-
ples of such algorithms. There are also options of the Wiener
Paper 12145 received Apr. 19, 2012; revised manuscript received Nov. 7, filter used in other than Fourier orthogonal transforms as,
2012; accepted for publication Nov. 12, 2012; published online Dec. 13, e.g., wavelet,16,19–21 DCT,4,22,23 and others.22 An attempt to
2012.
implement a nonlocal Wiener filter in spatial domain using
0091-3286/2012/$25.00 © 2012 SPIE and IS&T image “photometric similarities” is presented in Ref. 24.

Journal of Electronic Imaging 043020-1 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Reference 22 compares the Wiener-based filtering effi- auto-correlation function of the 2-D signal sðx; yÞ, and rn
ciency for different orthogonal bases. Although this is done is an auto-correlation function of the noise. Using the Fourier
for the 1-D case, an important conclusion is that the DCT transform property for convolution and the Wiener-Khinchin
domain Wiener filtering approaches the best known optimal theorem that relays correlation and power spectrum, one can
Karhunen-Loeve transform basis. This is due to very good obtain the Wiener-Hopf equation in the spectral domain
data de-correlation and the energy compaction properties of given for the 2-D case as:
the DCT, which are widely exploited in image and video
compression.25 Efficiency and usefulness of the local DCT ½Ps ðωx ;ωy Þ þ Pn ðωx ; ωy Þ
· H W ðωx ; ωy Þ ¼ Pus ðωx ; ωy Þ; (4)
commonly carried out in 8 × 8 pixel blocks has also been
proven for image denoising applications in Refs. 26–31. where Ps ðωx ; ωy Þ ¼ jF frs gj2 , Pn ðωx ; ωy Þ ¼ jF frn gj are
Thus, below we focus just on DCT as the considered basic power spectral densities of the noise-free image and noise,
orthogonal transform. respectively; F f·g denotes Fourier transform; ωx ; ωy are
In this paper, our goal is to analyze the potential of the spatial frequencies; Pus ðωx ; ωy Þ ¼ jF fpgj2 is a cross spec-
DCT image filtering in detail including an ideal (hypotheti- trum between noisy image and noise-free image; and
cal) case of a priori known global and local power spectra H W ðωx ; ωy Þ is a 2-D frequency response of the Wiener filter.
and a more practical case when only information on noise When the noise is not correlated with the image, p ¼ rs and
statistics (variance) is available. Next, we determine the the following expression holds:
potential limits of the DCT-based filtering efficiency for
fully overlapping blocks of 4 × 4, 8 × 8, and 16 × 16 pixels Pus ðωx ; ωy Þ ¼ Ps ðωx ; ωy Þ: (5)
within the Wiener approach and compare them to the results
obtained by the Chatterjee’s approach7,24 for a wide set of Thus, the Wiener filter in the spectral domain can be formu-
standard test images. Also, we analyze the filtering efficiency lated as
of the proposed multiscale DCT-based filters and compare
them to the state-of-the-art BM3D filter. Ps ðωx ; ωy Þ
H W ðωx ; ωy Þ ¼ : (6)
The paper is organized as follows: the image Wiener fil- Ps ðωx ; ωy Þ þ Pn ðωx ; ωy Þ
tering principle is considered; a way on how it reduces to
hard switching filter is shown in Sec. 2. Details of multiscale In practice, the exact power spectral densities Ps ðωx ; ωy Þ;
DCT-based filtering are presented in Sec. 3. Numerical Pn ðωx ; ωy Þ are often unavailable. A more realistic case pre-
simulation results for two proposed multiscale filters in sumes the use of the estimates of spectral densities:
comparison to the best known ones are presented in Sec. 4,
providing wide opportunities for analysis and comparisons. P^ s ðωx ; ωy Þ
A brief discussion of what else can be done in DCT-based H
^ W ðωx ; ωy Þ ¼ ; (7)
filtering is presented in Sec. 5. Finally, the conclusions P^ s ðωx ; ωy Þ þ P^ n ðωx ; ωy Þ
follow.
where H ^ W ðωx ; ωy Þ is an estimate of the frequency response
2 Image Wiener Filtering in DCT Domain of the Wiener filter and P^ s ðωx ; ωy Þ; P^ n ðωx ; ωy Þ are power
Let us consider an additive observation equation (model) spectral density estimates of the noise-free image and noise,
respectively.
uðx; yÞ ¼ sðx; yÞ þ nðx; yÞ; (1) In the case of additive white Gaussian noise, the model for
noise power spectral density is given by:
where uðx; yÞ is an observed noisy image; x, y are Cartesian
coordinates; sðx; yÞ denotes a noise-free image; and nðx; yÞ P^ n ðωx ; ωy Þ ¼ cðωx ; ωy Þ · σ 2 ; (8)
is a white Gaussian noise not correlated with sðx; yÞ. The
problem is to find an estimate of the noise-free image s^ ðx; yÞ where σ 2 is noise variance, cðωx ; ωy Þ is proportional to the
such that it minimizes MSE Ef½sðx; yÞ − s^ ðx; yÞ
2 g, where image size, and cð0; 0Þ ¼ 0 because we assume the Gaussian
Ef·g denotes the expectation operator. noise to have zero mean. Thus, the Wiener filter formula
The optimal linear filter that minimizes the MSE is the transforms to
well-known Wiener filter.14 It is the solution of Wiener-Hopf
equations expressed in matrix form as14 P^ s ðωx ; ωy Þ
^ W ðωx ; ωy Þ ¼
H : (9)
P^ s ðωx ; ωy Þ þ cðωx ; ωy Þ · σ 2
Rw ¼ p; (2)

where R is an autocorrelation matrix of a noisy image, w is In our proposal, we use the cosine transform instead of the
a vector of Wiener filter impulse response coefficients, and p Fourier transform for spectrum calculation, i.e., P^ s ðωx ;ωy Þ ¼
is a vector of cross-correlation between the noisy and noise- ½Sðωx ;ωy Þ
2 , where Sðωx ; ωy Þ is the DCT of a noise-free
free images. Alternatively, the Wiener-Hopf equations can be image (or its fragment). Again, in practice the noise-free
represented as image is not accessible to obtain Sðωx ; ωy Þ. For this reason,
the estimate of image power spectral density, P^ s ðωx ; ωy Þ,
r  w ¼ p; (3)
should be calculated using an observed noisy image. There-
fore, the image data has to be prefiltered to obtain some
where r ¼ rs þ rn is a vector of noisy image uðx; yÞ auto-
correlation function in the case of the additive noise rough estimate of a noise-free image Sðω ^ x ; ωy Þ and then to
model [Eq. (1)],  denotes convolution operation, rs is an calculate P^ s ðωx ; ωy Þ to implement the Wiener filter [Eq. (9)].

Journal of Electronic Imaging 043020-2 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

The last expression for the Wiener filter frequency are achieved [or, equivalently, in the sense of the peak signal-
response, Eq. (9), could be simplified assigning the unit gain to-noise ratio defined for byte represented images as
for all spatial frequencies where jUðωx ; ωy Þj ≥ βσ and PSNR ¼ 10 log10 ð65025∕MSEÞ]. This way, one can use
zero gain otherwise. This results in a hard thresholding local spectral estimates P^ s to take into account local data
technique5: activity for better noise filtering. For this purpose, the filter-

ing may be performed within blocks of m × m pixels, and
1 if jUðωx ; ωy Þj ≥ βσ such blocks are allowed to be overlapped for better noise
H T ðωx ; ωy Þ ¼ ; (10)
0 otherwise suppression. In this paper, we assume that the blocks are
maximally (fully) overlapped, i.e., the m × m neighboring
where β is a control parameter. If Sðωx ; ωy Þ is available, the blocks have the overlapping area of ðm − 1Þ × m pixels if
decision rule can be interpreted as jSðωx ; ωy Þj ≥ βσ, β ¼ 1 their upper left corner positions are shifted with respect to
that correspond to the Wiener filter pass band cutoff at the each other by only one pixel. In Refs. 23 and 26, it was
level of −3 dB. In practice, the decision rule is based on the shown that the DCT-based filtering with block overlapping
observed image, jUðωx ; ωy Þj ≥ βσ. reduces blocking effects and produces better output PSNR.
In this case, β was proven to have quasi-optimal value The DCT-based denoising with full overlapping is more
β ≈ 2.7.6,23,26 To confirm this, let us present some results. efficient in the sense of output MSE criterion than process-
Figure 1(a) shows a three-component LandsatTM image ing with partial overlapping or in nonoverlapped blocks.23
(optical bands) in red-green-blue representation. AWGN has Meanwhile, denoising in fully overlapped blocks takes
been added to all three components and they have been pro- more time. However, since DCT can be easily implemented
cessed by the DCT filter component-wise (8 × 8 pixel blocks using fast algorithms and/or specialized software or hard-
with full overlapping of blocks, see details in the next sec- ware, DCT-based denoising in fully overlapped blocks is
tions). The dependences of the output MSE for all three com- fast enough.
ponents are presented in Fig. 1(b) and 1(c) for noise standard So, for a locally adaptive Wiener DCT-based image filter
deviations 7 and 10, respectively. There are obvious minima we use a normalized DCT-2 transform32 given by
for all dependences for β slightly larger than 2.5. Since com-
ponent images are quite similar (characterized by cross- αðpÞαðqÞ X m−1 X
m−1

correlation factor of about 0.9), all dependences are very U ðmÞ ðp; qÞ ¼ uði þ k; j þ lÞ
m
similar. A general tendency is that optimal β shifts to larger 
k¼0 l¼0
   (11)
values for less complex images and/or larger standard devia- ð2k þ 1Þpπ ð2l þ 1Þqπ
tions of the noise and vice versa. Meanwhile, setting β equal × cos cos ;
2m 2m
to 2 or, e.g., 3.4 (i.e., 2.7  0.7) instead of 2.7 leads to an
MSE increase by about 10%. Thus, optimal setting (which where m × m is the block size; i, j are left upper corner coor-
is individual for each image and noise standard deviation) dinates of the data block in the full image;
instead of the recommended quasi-optimal is able to produce


output MSE which is only a few percent smaller than 1; 1 ≤ x ≤ m − 1
β ≈ 2.7. αðxÞ ¼ p1ffiffi ; x¼0 :
2
The thresholding filter [Eq. (10)] can be used as a preli-
minary image estimate s^ ðx; yÞ for its further use to determine
The inverse transform is given by
^ x ; ωy Þ for the Wiener filter [Eq. (9)].
Sðω
1Xm−1 X
m−1
uði þ k; j þ lÞ ¼ αðpÞαðkÞUðmÞ ðp; qÞ
3 Locally Adaptive Wiener Image Filter in DCT m p¼0 q¼0
Domain     (12)
ð2i þ 1Þkπ ð2j þ 1Þlπ
More accurate estimates of P^ s ðωx ; ωy Þ are used for Wiener × cos cos :
filtering, and better results in the sense of the output MSE 2m 2m

34 65

32
60
30
55
28
50
MSE

MSE

26
45
24
40
22

20 35

18 30
2 2.5 3 3.5 4 4.5 2 2.5 3 3.5 4 4.5
β β
(a) (b) (c)

Fig. 1 (a) Considered three-component image; (b) and (c) dependences of the output MSE on β.

Journal of Electronic Imaging 043020-3 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Using the definition in Eq. (11), the frequency response of filtered data that has to be aggregated to produce the filtered
the local hard thresholding filter is: image s^ ði; jÞ. The aggregation can be performed by aver-

aging the block pixels where the overlapping occurs. It can
ðmÞ 1 if jUðmÞ ðp; qÞj ≥ βσ also be performed using some weighting as proposed in
H T ðp; qÞ ¼ : (13)
0 otherwise Ref. 14, or using weighted least square patch averaging.
However, we have determined by simulations that this
The filtered image block is then obtained taking the inverse simple mean calculation for block data aggregation
transform as
Qði;jÞ
X s^ ðmÞ ði; j; qÞ
ðmÞ 1Xm−1 Xm−1
s^ ði; jÞ ¼ local
(18)
s^ T ði þ k; j þ lÞ ¼ αðpÞαðkÞU ðmÞ ðp; qÞ ðmÞ
m p¼0 q¼0 q¼1 Q ði; jÞ
 
ðmÞ ð2i þ 1Þkπ produces appropriately good results where s^ local ði; j; qÞ are
ðmÞ
× H T ðp; qÞ cos
2m i; j’th pixel of q’th overlapped block in Eq. (14) or Eq. (17)
  of size m, QðmÞ ði; jÞ denotes the number of overlapping
ð2j þ 1Þlπ
× cos : ð14Þ blocks in the i, j’th pixel. Note that filtering efficiency
2m
might be slightly worse for pixels near image edges since
for these pixels a smaller number of filtered values from pro-
Note that, opposite to scanning window filtering, the filtered
cessed overlapped blocks is aggregated (for example, only
values are obtained simultaneously for all pixels of a given
one for four image corner pixels).
block. And then, if processing with block overlapping is
Next, we have found by simulations that the aggregation
applied, these filtered values must be aggregated as described
of the overlapped blocks of different size might further
below.
improve noise suppression. To this end, at each pixel posi-
Next, we propose to use the estimate in Eq. (14) to deter-
tion, different values of m in Eqs. (11), (12), (14), and (17)
mine the local power spectrum P^ s ðp; qÞ as
are used and then the processed overlapped blocks of differ-

ent size are aggregated using some weighting. In particular,
^PðmÞ αðpÞαðqÞ X m−1 X
m−1
ðmÞ
s ðp;qÞ ¼ s^ T ½i þ k;j þ l
we have determined that the following weighting produces
m k¼0 l¼0 good results for different images and different noise levels:
   
ð2k þ 1Þpπ ð2l þ 1Þqπ 2 Qði;jÞ
× cos cos : ð15Þ X ð4Þ ð8Þ ð16Þ
0.15^slocal ði;j;qÞþ s^ local ði;j;qÞþ0.5^slocal ði;j;qÞ
2m 2m s^ ði;jÞ ¼ ;
q¼1 0.15Qð4Þ ði;jÞþQð8Þ ði;jÞþ0.5Qð16Þ ði;jÞ
Using Eq. (15), the frequency response of the local Wiener (19)
DCT-based image filter can be formulated as
where QðmÞ ði; jÞ is the number of overlapped blocks of size
^ ðmÞ P^ ðmÞ
s ðp; qÞ m × m. This approach will be further denoted as a multiscale
H W ðp; qÞ ¼ ; (16)
P^ ðmÞ
s ðp; qÞ þ c
ðmÞ ðp; qÞ · σ 2 DCT-based filter (MDF). The recommended weight setting
in Eq. (19) is based on the results presented in the next
where section.


0; if p ¼ q ¼ 0
cðmÞ ðp; qÞ ¼ 1 : 4 Simulation Results
m otherwise
The simulations have been performed using a wide set of
The filtered image block is obtained taking the inverse trans- standard grayscale test images33 shown in Fig. 2, all of
form as size 512 × 512 pixels. This allows obtaining quite full ima-
gination on properties and performance of different filtering
ðmÞ 1X m−1 Xm−1
^ ðmÞ
s^ W ði þ k; j þ lÞ ¼ αðpÞαðkÞUðmÞ ðp; qÞH W
algorithms and approaches considered in this paper. Noise
m p¼0 q¼0 variance (standard deviation) has been varied in a very wide
    range as well. Despite the noise standard deviation values of
ð2i þ 1Þkπ ð2j þ 1Þlπ the order 20 : : : 35 for grayscale images of 8-bit representa-
× ðp; qÞ cos cos :
2m 2m tion it is almost impossible to meet, in practice, the corre-
(17) sponding data often presented in literature dealing with filter
efficiency analysis and comparisons.7,12,14 Thus, we have
On the other hand, with the overlapping of the filtered blocks decided to obtain and present such data for the considered
in Eq. (14), Eq. (17) results in a high redundancy of the techniques.

Fig. 2 Test images: Lena, Boats, F-16, Man, Stream & bridge, Aerial, Baboon, Sailboat, Elaine, Couple, Tiffany, and Peppers.

Journal of Electronic Imaging 043020-4 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 1 Performance (in terms of the output PSNR, in dB) of the Table 1 (Continued).
standard DCT-based filtering techniques [Eqs. (9) and (10)] and the
ideal Wiener filtering that all operate over entire image transformed
data. DCT hard Wiener Ideal Wiener
Image σ thresholding filtering filtering

DCT hard Wiener Ideal Wiener Man 20 25.756 25.993 31.228

Image σ thresholding filtering filtering
25 24.785 25.047 30.3

Lena 2 39.797 39.916 44.936

30 23.981 24.268 29.575
5 34.089 34.247 39.398
35 23.267 23.578 28.983
10 30.472 30.671 35.795
Stream & bridge 2 39.808 39.843 43.373
15 28.44 28.676 33.895
5 31.533 31.654 36.896
20 27.025 27.294 32.631
10 26.940 27.103 32.704
25 25.931 26.23 31.697
15 24.886 25.069 30.568
30 25.032 25.364 30.96
20 23.608 23.809 29.194
35 24.251 24.61 30.356
25 22.704 22.922 28.206

Boats 2 39.883 39.976 44.421 30 21.974 22.211 27.448

5 33.213 33.354 38.468 35 21.385 21.64 26.839

10 29.112 29.289 34.56

Aerial 2 39.789 39.858 43.801
15 26.974 27.179 32.514
5 32.385 32.508 37.489
20 25.57 25.801 31.167
10 27.898 28.053 33.297
25 24.506 24.76 30.18
15 25.601 25.776 31.082
30 23.631 23.908 29.411
20 24.069 24.262 29.617
35 22.929 23.228 28.786
25 22.933 23.145 28.541

F-16 2 40.309 40.421 45.08 30 22.053 22.282 27.7

5 34.242 34.39 39.353 35 21.317 21.562 27.016

10 30.205 30.389 35.491 Baboon 2 40.105 40.124 43.148

15 27.984 28.198 33.409 5 31.524 31.617 36.405

20 26.41 26.651 32.011 10 26.244 26.387 31.942

25 25.26 25.524 30.971 15 23.778 23.946 29.649

30 24.31 24.596 30.15 20 22.313 22.499 28.175

35 23.456 23.761 29.474 25 21.342 21.545 27.122

Man 2 39.497 39.585 44.175 30 20.623 20.841 26.319

5 32.729 32.877 38.24 35 20.058 20.291 25.682

10 28.943 29.126 34.445 Sailboat 2 39.479 39.566 44.088

15 27.038 27.248 32.498 5 32.724 32.868 38.093

Journal of Electronic Imaging 043020-5 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 1 (Continued). Table 1 (Continued).

DCT hard Wiener Ideal Wiener DCT hard Wiener Ideal Wiener
Image σ thresholding filtering filtering Image σ thresholding filtering filtering

Sailboat 10 28.889 29.061 34.208 Peppers 2 39.475 39.589 44.646

15 26.823 27.019 32.167 5 33.608 33.772 39.064

20 25.406 25.626 30.811 10 30.239 30.44 35.51

25 24.332 24.572 29.806 15 28.345 28.579 33.641

30 23.46 23.722 29.014 20 26.981 27.241 32.388

35 22.701 22.983 28.364 25 25.862 26.15 31.451

Elaine 2 39.499 39.627 44.959 30 24.925 25.241 30.706

35 24.104 24.445 30.089

5 34.165 34.339 39.636

10 31.139 31.356 36.325

15 29.333 29.591 34.604

4.1 DCT Domain Hard Thresholding and Wiener
Denoising
20 28 28.298 33.448 Let us start by applying filtering to the entire image: the DCT
hard thresholding [Eq. (13)], practical Wiener filtering [with
25 26.877 27.21 32.58 spectrum estimation from DCT filtered image; Eq. (16)],
and the ideal Wiener (when Ps , Pn are both known). The
30 25.98 26.35 31.885 obtained results are presented in Table 1.
As can be easily expected, the output PSNR decreases if
35 25.118 25.519 31.307 noise standard deviation becomes larger (this tendency is
observed for any filtering approach). However, output
Couple 2 39.499 39.571 43.864 PSNR values differ a lot. For example, for the noise standard
deviation equal to 10, the DCT-based filtering with hard
5 32.193 32.332 37.751 thresholding (the quasi-optimal β ≈ 2.7 has been used for
all images and values of noise standard deviation) produces
10 28.245 28.421 33.854 output PSNR ranging from 31.14 dB for the simple structure
Elaine image to 26.24 dB for the complex structure Baboon
15 26.375 26.575 31.869
image. Similarly, the output PSNR for the ideal Wiener filter
ranges from 36.33 to 31.94 dB (again, for the test images
20 25.142 25.367 30.585
Elaine and Baboon, respectively).
25 24.225 24.472 29.657
A more detailed analysis shows that the output PSNR
values for the ideal Wiener filter are usually by 3 : : : 7 dB
30 23.472 23.743 28.938 larger than for the DCT-based filter with hard thresholding.
The difference slightly increases if the noise standard devia-
35 22.838 23.136 28.358 tion becomes larger. The difference is smaller for the test
images with more complex structure such as Baboon and
Tiffany 2 39.443 39.553 44.626
Stream & bridge.
The two-stage procedure of practical Wiener filtering pro-
5 33.458 33.62 39.084 duces intermediate results which are considerably closer to
the outputs of the DCT-based filter with hard thresholding
10 30.288 30.49 35.637 than to the ideal Wiener filter. The resulting PSNR for the
practical Wiener filter can be up to 0.4 dB better than for
15 28.609 28.849 33.883 the DCT-based filtering with hard thresholding. This means
that the estimates of the power spectrum P^ s ðωx ; ωy Þ are not
20 27.394 27.67 32.737 accurate enough. Note that the largest improvement for the
practical Wiener filter occurs for the test images with quite
25 26.376 26.69 31.9 simple structure and if the noise variance is large.
30 25.532 25.879 31.244
4.2 Block-Based Denoising
35 24.765 25.14 30.709 As it has been mentioned in the Introduction, images are 2-D
nonstationary processes for which local spatial spectra

Journal of Electronic Imaging 043020-6 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

shapes differ considerably from spatial spectra shapes for the The ideal Wiener filter again produces the output PSNR
corresponding entire images. Although 8 × 8 blocks are values that are by 3 : : : 4 dB larger than those corresponding
usually employed in the DCT-based filtering, we have con- to practically implementable methods. Note that for the ideal
sidered the question of block size selection in more detail. Wiener filter the best results are produced for m ¼ 16 and the
For this purpose, the output PSNR values have been obtained PLO PSNR for m ¼ 16 can be by almost 0.8 dB better than
for three sizes of m, namely 4, 8, and 16 taking into account for m ¼ 8.
that in such cases the DCT-based filtering can be carried out It is interesting to compare these results (that can be con-
faster than for other block sizes (e.g., m ¼ 11) that are, in sidered as PLO PSNR) to the corresponding data produced
general, also possible. The obtained results are presented by the Chatterjee’s approach.7 Such comparisons can be
in Table 2. As before, the results are given for the DCT- easily made by considering, e.g., the data in the last (right-
based filtering with hard thresholding, the practical (two- most) two columns of Table 2 (the best attainable values
stage) Wiener filtering [Eq. (17)], and the ideal Wiener fil- of PLO PSNR are marked bold). The PLO PSNR for the
tering. Besides, we present results for the lower bound of Chatterjee’s approach can be by almost 5 dB better (this
filtering efficiency obtained according to Ref. 7 using the takes place for simple structure images corrupted by AWGN
software tool offered by the authors34 (according to the with small standard deviation). Meanwhile, for complex
recommendations in Ref. 7, the selected number of clusters structure images such as Baboon and Stream & bridge, the
c ¼ 5 with the patch size ps z ¼ 11). The following results PLO PSNR for the Chatterjee’s approach can be by almost
are expressed not in output MSE as it is produced by the 4 dB smaller than for the ideal Wiener filter. For images of
software but in terms of PSNR for the convenience of com- middle complexity (as, e.g., Boat), the Chatterjee’s approach
parisons. The same test image set is used and the AWGN produces larger PLO PSNR for small standard noise devia-
with the same values of the standard deviation have been tions than the ideal Wiener filter and vice versa. One possible
simulated. explanation of this effect can be that it is a more difficult
The first observation that follows from comparison of the task to find similar patches and to take advantages of non-
corresponding data in Tables 1 and 2 is that the image block- local processing for images of more complex structure and
wise filtering produces considerably better results than the under condition where noise is intensive (has large variance).
image filtering with DCT applied to the entire image. The The results presented in Table 2 also confirm one obser-
output values for the block-wise version of the DCT-based vation earlier emphasized in Ref. 9. The output PSNR for the
filtering with hard thresholding are by 3 : : : 4 dB better DCT-based filtering with hard thresholding is quite close to
the Chatterjee’s limit7 for the complex structure images cor-
than the entire image counterpart. This once more confirms
rupted by intensive noise (see, e.g., data for the test images
expedience of the image local processing approach (with
Baboon and Stream & bridge for the noise standard deviation
block overlapping). Similar observations hold for the prac-
equal to 10 and larger). The difference is smaller than 1 dB.
tical and ideal Wiener filters.
Meanwhile, there is room for efficiency improvement for
As is seen, the block size m × m has sufficient impact on
simpler structure images if the noise standard deviation is
the DCT-based filter performance. The results for m ¼ 4 are
not large.
worse than for m ¼ 8 or 16 in practically all cases. The only
exceptions are the results for the test image Stream & bridge
for small standard noise deviations where PSNR for m ¼ 4 is 4.3 Comparison to the State-of-the-Art
slightly better than for m ¼ 16. Meanwhile, the PSNR values It becomes interesting to compare the performance of the
for m ¼ 8 and m ¼ 16 usually do not differ a lot between proposed DCT-based filters, MDF, and two-stage Wiener
each other, and simulations for m ¼ 32 revealed the filtering MDF with the state-of-the-art BM3D filter. The data which
efficiency reduction in comparison to m ¼ 16. The general allows carrying out such comparison are represented in
tendency is the following: m ¼ 16 is a better choice if the Table 3. First of all, the presented PSNR values for a given
noise standard deviation is larger and a processed image has image and noise standard deviation are quite close (the best
a simpler structure. results are marked bold). They differ by not more than 1 dB
We use the terms “simple structure” and “complex struc- (this happens for simple-structure images corrupted by
ture” images. Intuitively these terms are clear where the latter AWGN with large variance values, see data for the image
relates to more textural images. Unfortunately, until now Lena, σ ¼ 35). The BM3D filter performs better for some
there is no commonly accepted metric for image complexity. test images while the two-stage Wiener filter is better for
The practical Wiener filter [Eq. (17)] again produces per- others. It is difficult to establish some obvious performance
formance improvement compared to the DCT-based proces- dependence of these filters on image complexity. For two
sing with hard thresholding. Due to applying the Wiener simple-structure images such as Lena and Elaine, BM3D
filter at the second stage, the output PSNR can be increased results are better for Lena and the two-stage Wiener pro-
by up to 0.5 dB. We would like to stress here that the prac- duces, on average, better results for Elaine. Similarly, for two
tical Wiener filtering can be performed in a pipeline manner, complex structure test images, Baboon and Stream & bridge,
where the second stage processing is applied when the neces- the two-stage Wiener filter is better for the test image Stream
sary output data of the DCT-based thresholding is obtained. & bridge and vice versa.
Thus, although computation expenses are increased for the Setting the weights in Eq. (19), we have taken into
proposed two-stage procedure compared to the standard account that DCT-based denoising with 8 × 8 blocks usually
DCT-based denoising, the two-stage filtering is still consid- produces not worse filtering than with 16 × 16 blocks but
erably faster than most efficient denoising techniques that fewer artifacts are observed in neighborhoods of high-
search for similar blocks (patches), and is usually time contrast edges and small-sized objects. In turn, denoising
consuming. in 4 × 4 block is less efficient than for larger sizes of blocks.

Journal of Electronic Imaging 043020-7 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 2 Output PSNR (in dB) of the DCT-based image filters [Eqs. (14), (17), and (18)] in comparison to the noise suppression bound calculated
according to Ref. 7 (5 clusters were used with the patch size 11).

DCT with hard thresholding Wiener filtering Ideal Wiener filtering

PSNR
Image σ m
4 m
8 m
16 m
4 m
8 m
16 m
4 m
8 m
16 bound7

Lena 2 43.196 43.329 43.327 43.379 43.478 43.483 47.225 47.687 47.778 52.346

5 38.299 38.501 38.446 38.326 38.534 38.465 42.041 42.787 42.923 45.267

10 34.956 35.39 35.372 34.959 35.489 35.474 38.22 39.334 39.552 40.561

15 32.89 33.501 33.52 32.885 33.677 33.706 35.915 37.37 37.691 38.063

20 31.352 32.114 32.164 31.331 32.353 32.424 34.211 35.976 36.406 36.402

25 30.1 31.004 31.094 30.065 31.301 31.421 32.839 34.88 35.422 35.179

30 29.042 30.088 30.216 28.99 30.431 30.602 31.683 33.97 34.623 34.222

35 28.106 29.29 29.467 28.042 29.678 29.909 30.68 33.201 33.961 33.441

Boats 2 42.764 43.02 43.025 42.942 43.134 43.14 46.255 46.636 46.63 49.616

5 36.904 37.085 36.981 36.962 37.16 37.072 40.901 41.469 41.46 42.523

10 33.377 33.543 33.368 33.432 33.646 33.461 37.059 37.864 37.901 37.741

15 31.332 31.576 31.387 31.417 31.748 31.546 34.81 35.85 35.952 35.190

20 29.851 30.18 30.004 29.94 30.402 30.208 33.183 34.448 34.625 33.498

25 28.667 29.099 28.946 28.756 29.36 29.18 31.891 33.369 33.623 32.255

30 27.658 28.221 28.099 27.755 28.51 28.359 30.81 32.487 32.82 31.285

35 26.756 27.479 27.396 26.864 27.793 27.681 29.876 31.738 32.149 30.497

F-16 2 44.357 44.523 44.458 44.47 44.611 44.558 47.914 48.374 48.378 49.815

5 39.246 39.358 39.178 39.264 39.446 39.271 42.549 43.242 43.246 42.924

10 35.45 35.676 35.442 35.461 35.857 35.631 38.521 39.566 39.625 38.300

15 33.121 33.497 33.271 33.14 33.745 33.53 36.084 37.454 37.599 35.823

20 31.418 31.937 31.744 31.441 32.239 32.056 34.29 35.953 36.193 34.161

25 30.064 30.732 30.583 30.088 31.077 30.939 32.853 34.778 35.119 32.925

30 28.928 29.753 29.641 28.953 30.133 30.044 31.648 33.806 34.249 31.949

35 27.95 28.916 28.862 27.967 29.336 29.305 30.607 32.973 33.518 31.146

Man 2 43.364 43.373 43.211 43.485 43.452 43.283 46.611 46.806 46.663 49.059

5 37.448 37.436 37.12 37.566 37.556 37.249 41.151 41.554 41.432 41.731

10 33.439 33.44 33.119 33.565 33.626 33.282 37.26 37.946 37.879 36.945

15 31.284 31.328 31.049 31.396 31.535 31.215 34.989 35.929 35.931 34.525

20 29.81 29.933 29.698 29.905 30.152 29.872 33.346 34.522 34.599 32.968

Journal of Electronic Imaging 043020-8 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 2 (Continued).

DCT with hard thresholding Wiener filtering Ideal Wiener filtering

PSNR
Image σ m
4 m
8 m
16 m
4 m
8 m
16 m
4 m
8 m
16 bound7

Man 25 28.704 28.906 28.707 28.767 29.141 28.898 32.041 33.436 33.591 31.844

30 27.79 28.102 27.929 27.825 28.357 28.145 30.95 32.548 32.783 30.973

35 26.981 27.431 27.292 27 27.712 27.538 30.01 31.795 32.111 30.269

Stream & bridge 2 42.489 42.544 42.472 42.625 42.6 42.519 44.923 45.017 44.952 44.448

5 35.489 35.518 35.368 35.671 35.647 35.493 39.044 39.298 39.262 36.914

10 30.774 30.794 30.637 30.999 31.004 30.828 35.056 35.51 35.521 31.899

15 28.399 28.426 28.295 28.614 28.634 28.466 32.841 33.456 33.509 29.421

20 26.928 26.945 26.845 27.112 27.134 26.987 31.291 32.049 32.143 27.885

25 25.897 25.911 25.837 26.049 26.084 25.959 30.09 30.98 31.116 26.813

30 25.099 25.136 25.077 25.228 25.298 25.19 29.104 30.119 30.298 26.008

35 24.443 24.53 24.478 24.552 24.687 24.587 28.266 29.398 29.62 25.371

Aerial 2 43.299 43.239 42.913 43.345 43.223 42.935 45.82 45.883 45.622 45.471

5 36.777 36.641 36.167 36.824 36.695 36.27 39.994 40.236 39.985 38.169

10 32.25 32.156 31.704 32.353 32.3 31.848 35.902 36.361 36.155 33.305

15 29.759 29.737 29.362 29.914 29.933 29.526 33.575 34.217 34.066 30.781

20 28.071 28.112 27.805 28.252 28.342 27.991 31.927 32.737 32.645 29.130

25 26.819 26.902 26.655 27.003 27.155 26.858 30.64 31.608 31.575 27.926

30 25.84 25.954 25.75 26.015 26.22 25.966 29.577 30.694 30.722 26.991

35 25.031 25.179 25.007 25.193 25.454 25.234 28.667 29.926 30.015 26.234

DCT thresholding Wiener filtering Ideal Wiener filtering

PSNR
Image σ m
4 m
8 m
16 m
4 m
8 m
16 m
4 m
8 m
16 bound7

Baboon 2 42.151 42.302 42.34 42.319 42.38 42.392 44.396 44.603 44.648 44.137

5 34.933 35.095 35.111 35.125 35.225 35.23 38.339 38.7 38.782 36.472

10 30.198 30.356 30.347 30.397 30.523 30.499 34.243 34.774 34.897 31.186

15 27.685 27.874 27.879 27.907 28.08 28.055 31.997 32.67 32.829 28.466

20 26.027 26.248 26.274 26.249 26.471 26.46 30.443 31.248 31.444 26.745

25 24.837 25.065 25.118 25.037 25.292 25.302 29.248 30.178 30.411 25.539

30 23.93 24.168 24.237 24.101 24.383 24.412 28.272 29.321 29.592 24.640

35 23.213 23.458 23.537 23.349 23.655 23.703 27.444 28.605 28.916 23.942

Journal of Electronic Imaging 043020-9 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 2 (Continued).

DCT thresholding Wiener filtering Ideal Wiener filtering

PSNR
Image σ m
4 m
8 m
16 m
4 m
8 m
16 m
4 m
8 m
16 bound7

Sailboat 2 42.596 42.824 42.882 42.805 42.936 42.958 45.8 46.083 46.112 46.616

5 36.16 36.302 36.291 36.296 36.45 36.469 40.305 40.795 40.831 39.417

10 32.568 32.631 32.462 32.632 32.714 32.541 36.469 37.172 37.222 34.794

15 30.656 30.782 30.573 30.717 30.907 30.672 34.261 35.151 35.227 32.439

20 29.269 29.486 29.281 29.339 29.659 29.428 32.678 33.748 33.861 30.897

25 28.167 28.461 28.276 28.234 28.681 28.473 31.427 32.671 32.827 29.761

30 27.225 27.609 27.457 27.295 27.872 27.696 30.386 31.795 31.998 28.868

35 26.379 26.894 26.761 26.464 27.189 27.039 29.489 31.052 31.306 28.136

Elaine 2 42.385 42.688 42.92 42.628 42.81 42.992 45.944 46.403 46.732 54.793

5 35.907 36.275 36.737 36.095 36.485 36.951 40.782 41.529 41.946 47.596

10 32.92 33.18 33.481 32.872 33.148 33.483 37.186 38.198 38.643 42.807

15 31.621 31.938 32.089 31.521 31.927 32.055 35.076 36.332 36.808 40.294

20 30.637 31.105 31.201 30.508 31.161 31.235 33.535 35.043 35.563 38.561

25 29.756 30.416 30.496 29.603 30.545 30.624 32.294 34.056 34.631 37.304

30 28.933 29.811 29.885 28.765 30.003 30.11 31.242 33.251 33.891 36.316

35 28.155 29.251 29.331 27.975 29.505 29.652 30.32 32.58 33.292 35.510

Couple 2 42.725 42.868 42.84 42.918 43.004 42.966 46.984 47.256 47.165 49.022

5 37.076 37.147 36.963 37.178 37.234 37.046 41.597 42.078 42.004 50.355

10 33.323 33.463 33.25 33.429 33.605 33.377 37.628 38.411 38.406 42.740

15 31.131 31.389 31.216 31.261 31.585 31.38 35.267 36.329 36.405 37.514

20 29.573 29.942 29.821 29.706 30.179 30.022 33.546 34.863 35.024 34.828

25 28.346 28.843 28.767 28.482 29.108 29 32.177 33.725 33.972 33.116

30 27.34 27.959 27.915 27.469 28.251 28.183 31.033 32.791 33.124 31.897

35 26.472 27.21 27.201 26.594 27.533 27.503 30.047 31.997 32.413 30.971

Tiffany 2 43.468 43.583 43.544 43.63 43.699 43.656 47.414 47.848 47.864 54.471

5 38.397 38.563 38.41 38.484 38.668 38.518 42.258 42.992 43.091 47.068

10 34.896 35.191 35.096 34.947 35.353 35.24 38.451 39.595 39.806 42.125

15 32.899 33.343 33.323 32.912 33.537 33.485 36.134 37.658 37.986 39.581

20 31.437 32.073 32.124 31.429 32.308 32.327 34.407 36.281 36.73 37.937

Journal of Electronic Imaging 043020-10 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 2 (Continued).

DCT thresholding Wiener filtering Ideal Wiener filtering

PSNR
Image σ m
4 m
8 m
16 m
4 m
8 m
16 m
4 m
8 m
16 bound7

Tiffany 25 30.257 31.091 31.199 30.231 31.379 31.472 33.007 35.197 35.772 36.750

30 29.235 30.27 30.438 29.197 30.615 30.793 31.821 34.295 34.997 35.833

35 28.329 29.562 29.781 28.275 29.962 30.216 30.786 33.516 34.345 35.091

Peppers 2 42.67 42.902 42.985 42.917 43.097 43.149 46.734 47.143 47.204 52.776

5 37.309 37.415 37.384 37.345 37.465 37.464 41.634 42.306 42.408 45.475

10 34.471 34.653 34.477 34.419 34.679 34.484 37.932 38.906 39.058 40.663

15 32.706 33.112 32.928 32.649 33.217 33.016 35.724 36.991 37.206 38.161

20 31.259 31.936 31.781 31.22 32.115 31.958 34.096 35.646 35.937 36.512

25 30.033 30.951 30.844 30.002 31.202 31.103 32.783 34.599 34.971 35.301

30 28.969 30.101 30.047 28.933 30.414 30.382 31.673 33.732 34.19 34.353

35 28.024 29.345 29.347 27.984 29.717 29.756 30.706 32.986 33.531 33.579

Also, note that the DCT-based processing in blocks of DCT coefficient absolute values do not exceed 2σ and
different size can be carried out in parallel that allows P2.7σ is the probability that DCT coefficient absolute values
diminishing processing time. are larger than 2.7σ. One more characteristic of filtering
Figure 3 illustrates filtering efficiency for a fragment of efficiency has been determined: the ratio MSEout ∕σ 2 ,
the test image “Lena.” As is seen, noise removal is efficient where MSEout is output MSE after denoising. The obtained
and edge/detail preservation is good for both output images. data are presented in Table 4. The test images are put in
Figure 4 presents an example of processing the test image such order that P2σ in the fourth column increases.
“Baboon” by the proposed Wiener filter in comparison to The first observation is that the probabilities P2σ and P2.7σ
the state-of-the art BM3D filter. The BM3D filter suppresses are highly correlated. If P2σ is smaller, then P2.7σ is usually
noise better in “flat” (homogeneous image) regions while the larger. The second observation is that the values P2σ are
proposed filter preserves better texture and details; the fil- smaller and P2.7σ are larger for more complex-structure
tered image in this case has a more natural appearance. images and smaller noise variance values. This is clear
since for more complex-structure images the DCT coeffi-
5 Discussion cients for noise-free image have wider distribution. The
It is worth briefly discussing here the mechanism of DCT- third observation is that βopt increases if image complexity
based denoising with hard thresholding. Noise is removed in reduces and/or noise variance becomes larger. βopt varies
DCT-components of a block for which jUðp; qÞ < βσj from 2.3 to 2.8 where for most typical practical situations
(although hard thresholding operation simultaneously intro- βopt is within the limits from 2.6 to 2.7.
duces distortions in the corresponding signal components). It seems that if P2.7σ is preliminary determined for a given
Meanwhile, noise is preserved in the components when image under a condition of exactly known noise variance, it
jUðp; qÞ ≥ βσj. Therefore, noise reduction should increase can prove more careful threshold setting for providing cer-
if the number of DCT coefficient with jUðp; qÞ < βσj is tain benefits of filtering efficiency. Such a strategy can be
larger. treated as image/variance adaptive threshold setting. How-
All simulation results presented above for the DCT- ever, in our opinion, the benefits of this strategy are too
based denoising have been obtained for hard thresholding small to use in practice. A more reasonable way seems to
with the fixed β ≈ 2.7 in Eq. (13). However, as has been use locally adaptive setting of the thresholds, but currently
mentioned above, such threshold setting is quasi-optimal. we are unable to propose an algorithm to do this.
Let us demonstrate this by several examples. We have The data presented in Table 4 show that for noisy images
selected eight test images of different complexity widely their complexity (or, more strictly saying, complexity of
used in image processing applications. For three values image denoising task) can be indirectly characterized by
of noise standard deviation (5, 10, 15), the optimal values the parameter P2.7σ. Filtering is more efficient (smaller
βopt that provide maximal output PSNR have been deter- MSEout ∕σ 2 are provided) if P2.7σ is smaller. Note that
mined. They are presented in Table 4. Besides, we have MSEout ∕σ 2 can vary from 0.78 (less than 1 dB increase of
determined two probabilities: P2.7σ is the probability that output PSNR compared to input PSNR) to 0.13 and even

Journal of Electronic Imaging 043020-11 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 3 Performance (PSNR, in dB) of the proposed image filters Table 3 (Continued).
[Eqs. (14), (17), and (19)] in comparison to the images filtered by
the state-of-the art BM3D filter.14
Wiener
MDF MDF
Wiener [Eqs. (14) [Eqs. (17)
MDF MDF Image σ and (19)] and (19)] BM3D
[Eqs. (14) [Eqs. (17) Man 20 29.896 30.067 30.589
Image σ and (19)] and (19)] BM3D

Lena 2 43.407 43.546 43.594 25 28.896 29.082 29.616

5 38.555 38.558 38.724 30 28.115 28.323 28.86

10 35.488 35.566 35.932 35 27.471 27.707 28.224

15 33.639 33.795 34.269

Stream & bridge 2 42.553 42.573 42.662
20 32.283 32.508 33.051
5 35.511 35.605 35.775
25 31.211 31.497 32.071
10 30.794 30.976 31.174
30 30.33 30.668 31.27
15 28.44 28.615 28.789
35 29.576 29.963 30.557
20 26.978 27.126 27.271

Boats 2 43.101 43.184 43.181 25 25.96 26.086 26.228

5 37.115 37.181 37.283 30 25.195 25.31 25.46

10 33.541 33.613 33.92 35 24.595 24.703 24.862

15 31.576 31.719 32.14

Aerial 2 43.123 43.08 43.465
20 30.195 30.388 30.882
5 36.458 36.504 37.008
25 29.135 29.361 29.909
10 31.992 32.112 32.521
30 28.282 28.534 29.117
15 29.62 29.777 30.058
35 27.571 27.844 28.431
20 28.039 28.224 28.405
F-16 2 44.267 44.347 44.619
25 26.867 27.074 27.181
5 39.016 39.091 39.527
30 25.946 26.167 26.211
10 35.37 35.524 36.112
35 25.192 25.423 25.326
15 33.257 33.472 34.12
Baboon 2 42.368 42.406 42.303
20 31.765 32.033 32.711
5 35.173 35.273 35.104
25 30.616 30.933 31.637
10 30.43 30.568 30.394
30 29.668 30.038 30.76
15 27.962 28.135 27.902
35 28.857 29.281 29.985
20 26.351 26.541 26.277
Man 2 43.357 43.4 43.605
25 25.186 25.378 25.115
5 37.34 37.443 37.816
30 24.3 24.482 24.226
10 33.346 33.503 33.981
35 23.597 23.766 23.391
15 31.261 31.426 31.929

Journal of Electronic Imaging 043020-12 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 3 (Continued). Table 3 (Continued).

Wiener Wiener
MDF MDF MDF MDF
[Eqs. (14) [Eqs. (17) [Eqs. (14) [Eqs. (17)
Image σ and (19)] and (19)] BM3D Image σ and (19)] and (19)] BM3D

Sailboat 2 42.935 42.99 42.839 Tiffany 30 30.523 30.84 30.653

5 36.4 36.555 36.375 35 29.86 30.246 29.903

10 32.628 32.687 32.708

Peppers 2 43.044 43.19 42.917
15 30.759 30.844 30.86
5 37.497 37.551 37.535
20 29.47 29.604 29.571
10 34.653 34.636 34.947
25 28.465 28.647 28.569
15 33.125 33.186 33.502
30 27.639 27.864 27.737
20 31.985 32.128 32.371
35 26.942 27.203 26.928
25 31.045 31.265 31.419
Elaine 2 42.927 42.974 42.726
30 30.239 30.531 30.576
5 36.678 36.895 36.372 35 29.531 29.89 29.795

10 33.464 33.425 33.352

15 32.131 32.061 32.143

less (about 9 dB and more increase). Thus, it seems possible
to predict MSEout ∕σ 2 (or, equivalently, MSEout for a priori
20 31.276 31.274 31.296 known σ 2 ) from analysis of P2.7σ with practically acceptable
degree of accuracy. This can be one possible direction of
25 30.591 30.676 30.585 future research. It can be also expected that the use of poly-
nomial threshold operators and other more sophisticated
30 29.997 30.168 29.949

35 29.457 29.712 29.337

Couple 2 43.462 43.531 42.939

5 37.974 38.032 37.325

10 34.223 34.352 33.794

15 32.086 32.268 31.759

20 30.595 30.823 30.322

25 29.444 29.719 29.188

30 28.512 28.827 28.244

35 27.737 28.079 27.42

Tiffany 2 43.643 43.737 43.669

5 38.567 38.658 38.854

10 35.244 35.377 35.671

15 33.451 33.601 33.846

Fig. 3 Filtering results for the test image “Lena” contaminated by

20 32.233 32.419 32.535
AWGN with σ ¼ 25: (a) a fragment of the original image; (b) a
noisy fragment; (c) the proposed MDF filter [Eqs. (14) and (19)] output;
25 31.296 31.542 31.524 and (d) the proposed Wiener MDF [Eqs. (17) and (19)] output. Some
blocking effects can be noted on Lena’s face in (c).

Journal of Electronic Imaging 043020-13 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

Table 4 (Continued).

Image σ βopt P 2σ P 2.7σ MSEout ∕σ 2

Man 10 2.51 0.536 0.059 0.29

Boats 10 2.56 0.538 0.058 0.29

F-16 10 2.69 0.557 0.046 0.19

Peppers 10 2.63 0.560 0.041 0.22

Man 15 2.57 0.561 0.041 0.21

Lena 10 2.7 0.561 0.042 0.19

Boats 15 2.61 0.561 0.042 0.2

Tiffany 10 2.6 0.566 0.037 0.2

F-16 15 2.74 0.572 0.035 0.14

Lena 15 2.8 0.575 0.032 0.13

Peppers 15 2.77 0.576 0.031 0.14

Tiffany 15 2.7 0.580 0.027 0.13

Fig. 4 Filtering results for the test image “Baboon” contaminated by

AWGN with σ ¼ 25: (a) a fragment of the original image; (b) a noisy
fragment; (c) the output of the BM3D filter; and (d) the proposed
thresholds35,36 can improve performance of the DCT-based
Wiener MDF [Eqs. (17) and (19)] output. The picture in (d) looks denoising.
more natural.
6 Conclusions
Different approaches to filtering grayscale images corrupted
Table 4 DCT-based filter efficiency and DCT coefficient statistics for by AWGN are considered including the DCT-based denois-
different test images and noise variances. ing with hard thresholding, two-stage Wiener filter, and ideal
Wiener filters that are compared to the state-of-the art BM3D
technique. Several sizes of fully overlapped image blocks
Image σ βopt P 2σ P 2.7σ MSEout ∕σ 2
are studied and it is shown that processing in 8 × 8 and
Baboon 5 2.3 0.340 0.233 0.78 16 × 16 pixel blocks produces approximately the same
results. It has been demonstrated that the performance can
Stream & bridge 5 2.38 0.369 0.204 0.71 be slightly improved by combining the filter outputs that
perform processing using different block sizes. Following
Baboon 10 2.34 0.450 0.128 0.58 this approach, two multiscale DCT-based filters, MDF and
Wiener MDF, are proposed and their properties analyzed.
Man 5 2.45 0.474 0.111 0.46 Potential limits of output PSNR (or MSE) for the ideal
Wiener filter and Chatterjee’s approach are obtained and
Stream & bridge 10 2.37 0.474 0.105 0.52 compared. These limits are, on average, of the same order but
can differ by up to 5 dB depending on the image processed
Boats 5 2.38 0.476 0.107 0.49 and noise variance. Thus, we can state that the potential
limits of filtering efficiency are “approach-dependent.”
Baboon 15 2.37 0.501 0.083 0.47 The state-of-the-art filters including the DCT-based
denoising and the Wiener-based techniques provide filtering
Peppers 5 2.35 0.509 0.076 0.45
performances quite close to Chatterjee’s limit for complex-
F-16 5 2.56 0.518 0.077 0.32
structure images and large noise variance. Performance char-
acteristics of the state-of-the art BM3D filter and the pro-
Lena 5 2.5 0.519 0.073 0.36 posed Wiener MDF are very close while the latter filter is
simpler and faster.
Stream & bridge 15 2.37 0.521 0.067 0.4 The proposed MDF techniques require less computational
time than the BM3D filter and, especially, the Chatterjee
Tiffany 5 2.49 0.523 0.069 0.36 filter, which requires image clustering to perform nonlocal
averaging. MDF technique [Eqs. (14) and (19)] is about
two times faster than the Wiener MDF [Eqs. (17) and (19)]

Journal of Electronic Imaging 043020-14 Oct–Dec 2012/Vol. 21(4)

 Pogrebnyak and Lukin: Wiener discrete cosine transform-based image filtering

and produces good visual quality of the filtered images when 21. S. Suhaila and T. Shimamura, “Image restoration based on edgemap
and Wiener filter for preserving fine details and edges,” Int. J. Circuits,
the noise variance is low (σ < 0.1). Syst. Signal Process. 5(6), 618–626 (2011).
It has also been shown that filtering efficiency depends 22. S. Rahardja and B. Falkowski, “Comparative study of discrete orthog-
considerably on DCT coefficient statistics. A more detailed onal transforms in adaptive signal processing,” IEICE Trans. Fundam.
E-85-A(8), 1386–1390 (1999).
study of this dependence can be a direction of future research 23. V. V. Lukin et al., “Image filtering based on discrete cosine transform,”
to further improve performance of the block-wise DCT- Telecommun. Radio Eng. 66(18), 1685–1701 (2007).
24. P. Chatterjee and P. Milanfar, “Patch-based near-optimal image denois-
based filters. ing,” IEEE Trans. Image Process. 21(4), 1635–1649 (2012).
25. A. Kaarna, “Compression of spectral images,” in Vision Systems: Seg-
Acknowledgments mentation and Pattern Recognition, G. Ohinata and A. Dutta, Eds.,
pp. 269–298, I-Tech, Vienna, Austria (2007).
We are thankful to anonymous reviewers for their valuable 26. R. Oktem et al., “Locally adaptive DCT filtering for signal-dependent
comments and propositions. This work was partially sup- noise removal,” EURASIP J. Adv. Signal Process. 2007, 042472
(2007).
ported by Instituto Politecnico Nacional as a part of the 27. V. Lukin et al., “Adaptive DCT-based filtering of images corrupted by
research project SIP20120530. spatially correlated noise,” Proc. SPIE 6812, 68120W (2008).
28. N. N. Ponomarenko et al., “3D DCT based filtering of color and multi-
References channel images,” Telecommun. Radio Eng. 67(15), 1369–1392 (2008).
29. V. Lukin et al., “Discrete cosine transform-based local adaptive filtering
1. W. K. Pratt, Digital Image Processing, 4th ed., Wiley-Interscience, of images corrupted by nonstationary noise,” J. Electron. Imaging
New York (2007). 19(2), 15 (2010).
2. A. Bovik, Handbook of Image and Video Processing, p. 1429, Aca- 30. K. Egiazarian et al., “Adaptive denoising and lossy compression of
demic Press, USA (2000). images in transform domain,” J. Electron. Imaging 8(3), 233–245
3. R. Touzi, “A review of speckle filtering in the context of estimation (1999).
theory,” IEEE Trans. Geosci. Remote Sens. 40(11), 2392–2404 (2002). 31. R. Oktem et al., “Transform based denoising algorithms: comparative
4. A. Foi, “Pointwise shape-adaptive DCT image filtering and signal- study,” J. Electron. Imaging 11(2), 149–156 (2002).
dependent noise estimation,” Ph.D. Thesis, p. 194, Tampere University 32. A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
of Technology, Tampere, Finland (2007). Processing, 2nd ed., Prentice Hall, New Jersey (2009).
5. K. N. Plataniotis and A. N. Venetsanopoulos, Color Image Processing 33. “The USC-SIPI image database, volume 3: miscellaneous,” http://sipi
and Applications, Springer-Verlag, New York, p. 355 (2000). .usc.edu/database/database.php?volume=misc (19 September 2012).
6. V. Lukin et al., “Adaptive DCT-based filtering of images corrupted by 34. P. Chatterjee and P. Milanfar, “Fundamental limits of image denoising,”
spatially correlated noise,” Proc. SPIE 6812, 68120W (2008). 09 February 2012, http://users.soe.ucsc.edu/~priyam/bounds.
7. P. Chatterjee and P. Milanfar, “Is denoising dead?” IEEE Trans. Image 35. C. Smith, S. Agaian, and D. Akopian, “A wavelet-denoising approach
Process. 19(4), 895–911 (2010). using polynomial threshold operators,” IEEE Signal Process. Lett.
8. X. Zhu and P. Milanfar, “Automatic parameter selection for denoising 15, 906–909 (2008).
algorithms using a no-reference measure of image content,” IEEE 36. V. Lukin, N. Ponomarenko, and K. Egiazarian, “HVS-metric-based per-
Trans. Image Process. 19(12), 3116–3132 (2010). formance analysis of image denoising algorithms,” in European Work-
9. V. Lukin et al., “Image filtering: potential efficiency and current pro- shop on Visual Information Processing (EUVIP), Paris, France,
blems,” in International Conf. on Acoustics, Speech and Signal Proces- pp. 156–161, IEEE (2011).
sing (ICASSP 2011), Prague, Czech Republic, p. 4, IEEE (2011).
10. D. Fevralev et al., “Efficiency analysis of color image filtering,”
EURASIP J. Adv. Signal Process. 2011(1), 1–19 (2011). Oleksiy Pogrebnyak received his PhD
11. M. Elad, Sparse and Redundant Representations. From Theory to degree from Kharkov Aviation Institute (now
Applications in Signal and Image Processing, p. 376, Springer National Aerospace University), Ukraine, in
Science+Business Media, Berlin (2010). 1991. Currently, he is with The Center for
12. C. Kervrann and J. Boulanger, “Local adaptivity to variable smoothness
for exemplar-based image regularization and representation,” Int. J. Computing Research of National Polytechnic
Comput. Vision 79(1), 45–69 (2008). Institute, Mexico. His research interests
13. K. Dabov et al., “Color image denoising via sparse 3D collaborative include digital signal/image filtering and
filtering with grouping constraint in luminance-chrominance space,” compression, and remote sensing.
in Proc. IEEE Int. Conf. on Image Process. (ICIP 2007), San Antonio,
Texas, pp. 313–316, IEEE (2007).
14. K. Dabov et al., “Image denoising by sparse 3-D transform-domain col-
laborative filtering,” IEEE Trans. Image Process. 16(8), 2080–2095
(2007).
15. S. Haykin, Adaptive Filter Theory, 4th ed., Prentice Hall Int. Inc., Vladimir V. Lukin graduated from Kharkov
New Jersey (2002). Aviation Institute (now National Aerospace
16. F. Jin et al., “Adaptive Wiener filtering of noisy images and image
sequences,” in IEEE Int. Conf. Image Process. III (ICIP 2003), Barce- University) in 1983 and got his diploma
lona, Spain, pp. 349–352, IEEE (2003). with honors in radio engineering. Since
17. J. S. Lee, “Digital image enhancement and noise filtering by use of then he has been with the Department of
local statistics,” IEEE Trans. Pattern Anal. Mach. Intell. PAMI-2(2), Transmitters, Receivers and Signal Proces-
165–168 (1980). sing of National Aerospace University. He
18. D. T. Kuan et al., “Adaptive noise smoothing filter for images with defended the thesis of Candidate of Techni-
signal dependent noise,” IEEE Trans. Pattern Anal Mach. Intell. cal Science in 1988 and Doctor of Technical
PAMI-7(2), 165–177 (1985). Science in 2002 in DSP for Remote Sensing.
19. G. M. Rajathi and R. Rangarajan, “Efficient adaptive Wiener filter with Since 1995 he has been in cooperation with
thresholding for better image enhancement,” Eur. J. Sci. Res. 69(1),
143–153 (2012). Tampere University of Technology. Currently, he is department vice
20. P.-L. Shui and Y.-B. Zhao, “Image denoising algorithm using doubly chairman and professor. His research interests include digital signal/
local Wiener filtering with block-adaptive windows in wavelet domain,” image processing, remote sensing data processing, image filtering,
J. Signal Process. 87(7), 1721–1734 (2007). and compression.

Journal of Electronic Imaging 043020-15 Oct–Dec 2012/Vol. 21(4)