Está en la página 1de 10

IET Image Processing

Research Article

Curvelet-based multiscale denoising using ISSN 1751-9659


Received on 5th August 2017
Revised 3rd December 2017
non-local means & guided image filter Accepted on 27th December 2017
E-First on 31st January 2018
doi: 10.1049/iet-ipr.2017.0825
www.ietdl.org

Susant Kumar Panigrahi1 , Supratim Gupta1, Prasanna K. Sahu1


1Department of Electrical Engineering, National Institute of Technology, Rourkela-769008, India
E-mail: susant146@gmail.com

Abstract: This study presents an image denoising technique using multiscale non-local means (NLM) filtering combined with
hard thresholding in curvelet domain. The inevitable ringing artefacts in the reconstructed image – due to thresholding – is
further processed using a guided image filter for better preservation of local structures like edges, textures and small details.
The authors decomposed the image into three different curvelet scales including the approximation and the fine scale. The low-
frequency noise in the approximation sub-band and the edges with small textural details in the fine scale are processed
independently using a multiscale NLM filter. On the other hand, the hard thresholding in the remaining coarser scale is applied
to separate the signal and the noise subspace. Experimental results on both greyscale and colour images indicate that the
proposed approach is competitive at lower noise strength with respect to peak signal to noise ratio and structural similarity index
measure and excels in performance at higher noise strength compared with several state-of-the-art algorithms.

1 Introduction key signal components of the image by thresholding out the noise.
The universal VisuShrink [5], the unbiased estimator-based
The restoration of the original image from its noisy observation – SureShrink [6], the Bayesian prior-based BayesShrink [7], the
contaminated during acquisition, reception or transmission – is a statistical co-dependent bivariate shrinkage-based BiShrink [8], the
stepping stone for many image processing or computer vision probabilistic shrinkage function-based ProbShrink [9], the linear
tasks. Often the characteristics of noise depend on the sensor type, weight estimation-based Stein's unbiased risk estimate-linear
pixel dimension, exposure time, ISO speed, temperature, and expansion of thresholds (SURE-LET) [10] and the recently
ambient illumination level. However, in most of the applications, proposed NeighShrink [11] are few thresholding techniques (λγ in
the distribution of noise can be characterised as additive white
(2b)) that were applied in wavelet domain to preserve the high
Gaussian noise (AWGN)
magnitude signal coefficients. However, the inability of wavelets in
y = z + η, (1) representing the curved edges (C2 singularity) degrades the
performance of wavelet-based image denoising approaches. The
where y is the observed (noisy) image, z is the latent image and implementation of anisotropic scaling in curvelet transform aides
η ∈ ℵ(0, σ 2) is the Gaussian noise of zero mean and σ 2 variance. in representing image edges more sparsely compared with
wavelets. Starck et al. [12] employed hard thresholding on the
Available denoising methods can be broadly categorised into
magnitude of complex curvelet coefficients for denoising.
three classes: spatial, transform and dual domain approaches. The
However, these transform domain-based thresholding (shrinkage)
underlying idea behind spatial filtering techniques differs only to
techniques suffer from inevitable ringing artefacts due to a sudden
the extent of the weights (kernels) that are calculated – either
jump in coefficient magnitudes.
locally or non-locally – to estimate different data points in an
Dual domain or hybrid approaches combine the advantages of
image [1]. Bilateral filter (BF) [2], non-local means filter (NLM)
both spatial and transform domain techniques to improve the
[3] and recently proposed guided image filter (GIF) [4] are few
overall image denoising quality. The multi-resolution BF (MBF)
modern edge preserving filters that exploit either local, non-local
incorporated both BF and wavelet thresholding in the
or both self-similarity among the image patches for its restoration
approximation and the detail scales to suppress simultaneously the
in the spatial domain. In contrast to spatial domain approaches,
coarser grain (or low-frequency) and fine grain (or high-frequency)
transform domain techniques represent signals with fewer non-zero
noise [13]. Knaus et al. [14] proposed a dual domain image
coefficients. A threshold, as given in (2b), is employed in this
denoising method by integrating joint BF and short time Fourier
approach to separate the signal and the noise subspace for image
transform-based wavelet shrinkage technique. The inability of
restoration
wavelets in representing edges [15] and the limitation of BF in
preserving gradient direction of edges [4] may degrade the overall
Y γ = Tγ y , (2a)
performance of such techniques. Recently, NLM filter – instead of
BF – has been considered in several kinds of literature to improve
^ Y γ, if | Y γ | > lγ, the performance of hybrid domain approaches [16–18]. However,
Yγ = (2b)
0, otherwise, the block matching 3D collaborative filter (BM3D) excelled in
denoising by grouping the similar (non-local) patches and
^
z^ = Tγ−1 Y γ . (2c) collaboratively filtering the 3D blocks using 1D wavelet
thresholding [19]. Though BM3D is considered the state-of-the-art
technique, it is still unable to denoise few homogeneous regions
Here, the transformed coefficients Y is thresholded at each scale γ
that manifest as low-frequency noise [20]. Moreover, the patch-
and reconstructed using inverse transform T −1 to generate the based methods that search for more number of patches with similar
denoised image z^. local spatial structures may reduce at a higher noise level and thus
The energy-compaction property of several multi-resolution restricts the performance of denoising. Motivated by the fact that
signal transformation T is proved to be suitable for preserving the the residual sparsity among the non-local similar patches can be

IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918 909


© The Institution of Engineering and Technology 2018
reduced under a constrained prior model, recently Zha et al. [21] The weight represents the similarity between two image patches,
proposed a new image denoising technique – called group sparsity N(p) and N(q), centred at two pixels p and q and it is defined as
residual constrain (GSRC) – to enhance the performance of group-
sparse-based methods (that includes state-of-the-art BM3D ∥ y(N(p)) − y(N(q)) ∥22, a
technique). Though the GSRC technique achieves numerically W(p, q) = exp − , (4)
h2
improved results with a modest increase in visual quality at higher
noise strength, the iterative method puts higher computational
where ∥ ⋅ ∥2,2 a is the Gaussian modulated Euclidean distance
burden. Similarly, the sophisticated Cauchy filter-based image
denoising technique provides comparable results with higher between two patches and a is the standard deviation of the
computation complexity [22]. Gaussian kernel. The parameter h = kσ controls the smoothness of
The NLM – an averaging filter – utilises the similarity among the denoised image.
the patches to denoise the image. Though it is very efficient in Though similar patches are searched in a larger neighbourhood,
preserving edges, it is unable to process efficiently near the it is still local compared with the whole image [23]. Moreover, the
textured regions. The energy compaction property of the curvelet patches of fixed square shape and a fixed scale over the
that represents any signal in several scales can be used to separate neighbourhood may limit the performance of the NLM filter [24].
various spatial frequencies of image efficiently. We have Unlike spatial domain implementation [3], the multiscale NLM
incorporated the advantages of both (multiscale) NLM filtering and filter has the potential to suppress the noise in the approximation
hard thresholding in three different scales of the curvelet: the and in the finest scale by exploiting the advantages of intra-scale
approximation, the coarser and the fine scale. The edge preserving similarities among the curvelet coefficients [25]. Furthermore, the
the property of the non-linear NLM filter ensures the suppression high energy compactness of these coefficients also aids in the more
of noise in the approximation scale and aids in preserving well- accurate estimation of NLM weights in the curvelet domain.
connected edges with small image details in the fine scale. Unlike
multiscale filtering, the hard thresholding in the coarser scale – at 2.2 Curvelet thresholding (CT)
different orientations – is employed to separate the signal from the
The ‘sparse-land’ modelling represents an image with a few non-
insignificant noise coefficients. The inevitable ringing artefacts in
zero coefficients of high magnitude in different resolutions and
the reconstructed image are further processed by GIF to obtain the
directions. Threshold, λγ, o at any scale, γ is derived to separate the
final denoised image. We extend our single channel algorithm to
multichannel colour image denoising using luminance high magnitude signal coefficients from the insignificant noise
chrominance colour channels (YUV) colour space transformation. component [12]
The performance of the proposed hybrid technique is tested on
both greyscale and colour images. The experimental results – λγ, o = kσσγ, o, (5)
compared with several state-of-the-art techniques – demonstrate
the competitiveness of the proposed approach at lower noise where k is a scale-dependent constant and the variance σγ2, o is
strength, while yielding better performance at a higher value of σ . estimated for each scale, γ and orientation, o using Monte-Carlo
The rest of the article is organised as follows: we first review simulations of the Curvelet transform of a few standard white noise
few works related to our methodology in Section 2. The proposed images [12].
algorithms for both greyscale and RGB image are explained in As illustrated in Fig. 1, the threshold (as defined in (5)) removes
Section 3. The experimental results in terms of denoising quality the noise as well as a few fine details of the image. Moreover, the
and computational complexity are presented and discussed in ringing artefacts deteriorate the localisation of edges in the
Section 4. Finally, Section 5 concludes the paper. reconstructed image. To minimise the effect, we restrict ourselves
to three decomposition levels. The threshold is then applied in the
2 Background methods coarser scale to separate the insignificant noise coefficients from
the signal. The ringing artefacts in the reconstructed image is
Often, the frequency characteristics of AWGN are neglected in further processed using GIF.
many image denoising approaches. The noise can be as smooth as
an image or as oscillatory as many fine structures in the image. 2.3 Guided image filter
Multi-resolution techniques have proven particularly well in
separating signal and noise subspace better at one resolution level GIF is a fast [O(N) for an image with N pixels.], edge-preserving,
than another. However, simply the hard or soft thresholding at locally optimised, linear translation invariant filter [4]. The
different resolution levels proved to be inadequate in distinguishing optimisation of such a filter consists of two terms: a data term and
the fine texture details and the noise. Here, the combined approach a regularisation term. The data term measures the fidelity between
of multiscale NLM filter and hard thresholding in curvelet domain the restored image and the image to be filtered, while the
is considered to remove the noise in more than one resolution level. regularisation term provides the smoothness level of the
We briefly reviewed these methodologies along with GIF – which reconstructed image. Assuming y(p) to be the noisy image at any
is employed to suppress the ringing artefact – in the following given pixel p, then it can be decomposed into two parts called the
subsections. base layer z^(p) and the detail layer e(p)

2.1 NLM filter y(p) = z^(p) + e(p) . (6)


NLM is a non-linear weighted average filter. The weights are A linear transform of an external image with better detail
selected using a ‘metric’ that non-locally determines whether the information about the input image known as the guidance image, G
pixels are similar or not [3]. Assuming any pixel p with its intensity is considered in the formulation of GIF to filter the noisy image y
denoted by y(p), the NLM estimation, NLM [y](p) of the noisy as [4]
image y is computed as [3]
z^(p) = ap′G(p) + bp′, ∀p ∈ wς(p′), (7)
1
C(p) q∑
NLM[y](p) = W(p, q)y(q), (3)
∈Ω where the linear parameters ap′ and bp′ are defined inside a small
window, wς(p′) centred at p′ with radius ς. These parameters
where the weight W(p,q) is searched within a neighbouring ensure the preservation of edge in z^ if and only if G has an edge,
window Ω such that i.e. ∇z^ = ap′∇G.
1
C(p) q∑
W(p, q) = 1.
∈Ω

910 IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918


© The Institution of Engineering and Technology 2018
Fig. 1  Curvelet coefficients of
(a) Clean Lena image 4, (b) Noisy image σ = 25 , (c) Thresholded image

Fig. 2  Performance of CT technique – with NLM filtering in the approximation & the fine scale – in terms of
(a) PSNR, (b) SSIM measure for a number of decomposition levels Nγ

To determine the linear coefficients ap′, bp′ a constraint (as Therefore, the filtered image z^ in (11) is obtained as the local
given in (6)) is added to minimise the cost function E ap′, bp′ , combination of some percentage of the input image with its local
which is defined as mean. In addition, the following observations may be noted:

• Case 1: ϵ ≫ σG2 , ς(p′) : ap′ ≃ 1&bp′ ≃ 0 : GIF preserves the


E= ∑ (ap′G(p) + bp′ − yi)2 + ϵa2p′ . (8)
p ∈ wς(p′) details as it encounters ‘edge’.
• Case 2: ϵ ≪ σG2 , ς(p′) : ap′ ≃ 0&bp′ ≃ 1 : GIF filters the image as
Here ϵ denotes the regularisation parameter [the regularisation it encounters a ‘homogeneous’ region.
parameter, ϵ = k3σ is related to noise standard deviation σ to serve
as a tuning parameter for better edge localisation in the proposed 3 Proposed denoising framework
algorithm] that penalises large ap′. The optimisation of (8) provides
the solution for the linear coefficients ap′, bp′ [4] The NLM filter requires the noise to remain additive in the
analysing domain. The use of tight frames to represent any square-
1 integrable function f in Curvelet domain obeys Parseval's identity
|w| ∑ p ∈ wς( p′)
G(p)y(p) − μG, ς(p′)μy, ς(p′) as [15]
ap′ = , (9)
σG2 , ς(p′) + ϵ
f = ∑ ⟨ f , ϕγ τ o⟩ϕγ τ o,
, , , , (13)
γ, τ, o
bp′ = μy, ς(p′) − ap′ μG, ς(p′), (10)
where ϕγ, τ, o denotes the curvelet basis function. The linearity
where μG, ς(p′) and σG2 , ς(p′) are the mean and variance of G in property of the curvelet motivated the authors to implement a
wς(p′). Similarly, μy, ς(p′) denotes the mean of the noisy image y. (multiscale) NLM filter in the approximation and the fine scale. We
The final computation of z^(p) assumes the linear coefficients to carried out experiments to obtain an optimum number of
be constant in wς(p′) such that decomposition levels Nγ for image restoration using CT under
maximum peak signal to noise ratio (PSNR) and structural
z^(p) = ā pG(p) + b̄ p, (11) similarity index measure (SSIM). Assuming the approximation and
the fine scale coefficients are filtered via NLM and the coarser
where ā p = (1/( | wς(p′) | ))∑ p′ ∈ wς( p) ap′ and b̄ p = scale(s) coefficients are thresholded, the PSNR and SSIM measure
for an image decomposed in four different scales, Nγ = 3, 4, 5, 6 are
(1/ | wς(p′) | )∑ p′ ∈ wς(p) bp′ are the average values of ap′ and bp′ for
calculated. The mean and the standard deviation plot of these
all the overlapping windows that cover the pixel p. indices for Cameraman image (of size 512 × 512) contaminated
In the proposed algorithm, GIF is used as an ‘edge-preserving’ with AWGN of σ between 10 and 75 is shown in Fig. 2. It is
filter to suppress the ringing artefacts near edges. Thus, (9) & (10) observed that the minimum decomposition levels, Nγ = 3 with only
being reformulated by assuming G ≡ y, such that
coarser scale coefficients being thresholded may enhance the
quality of the denoised image.
σy2, ς(p′) The ringing artefacts due to hard thresholding in the coarser
ap′ = , (12a)
σy2, ς(p′) + ϵ scale are further suppressed by incorporating GIF in the
reconstructed image, as shown in Fig. 3. We extend the greyscale
bp′ = (1 − ap′)μy, ς(p′) . (12b) (single channel) algorithm to multichannel (colour) image
restoration by independently applying it to a ‘decorrelated’
luminance/colour-difference space (YUV) as discussed in Section
3.2.

IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918 911


© The Institution of Engineering and Technology 2018
Fig. 3  Illustration of proposed denoising algorithm for greyscale image

Fig. 4  Curvelet based reconstruction from the approximation coefficients of


(a) Original Lena image, (b) Noisy σ = 40 image, (c) NLM-based filtered image

Fig. 5  Curvelet coefficients in the fine scale for


(a) Original texture image, (b) Recovered details from the AWGN of σ = 25 using CT, (c) Recovered details from the AWGN of σ = 25 using NLM-based filtering

due to NLM filtering in Fig. 5c. However, the hard thresholding


[12] removes noise as well as many small details of the image in
the fine scale as shown in Fig. 5b. The disconnected edges may
introduce granular artefacts in the final denoised image.
The distortion near edges as shown in Fig. 6a degrade the
overall denoising quality of the reconstructed image. The ‘edge-
aware’ GIF is considered as a post-possessing mechanism to
suppress the ringing artefacts and to preserve the local structures
Fig. 6  Proposed denoising method for Lena image σ = 40 like edges, textures and small details in the restored image. The
(a) Before (PSNR = 29.8557, SSIM = 0.8438) the application of GIF, (b) After (PSNR  visual improvement in Fig. 6 and the quantitative improvement in
= 30.2377, SSIM = 0.8752) the application of GIF PSNR and SSIM measures validate the consideration of GIF as an
edge preserving filter in the proposed algorithm.
3.1 Greyscale image denoising
As illustrated in the block diagram (Fig. 3), the image is initially 3.2 Colour image denoising
decomposed in three different scales in order to consider both the The most straightforward way of multichannel image denoising is
low-frequency or coarser grain and high-frequency fluctuations or to employ the greyscale algorithm for each colour channel (RGB),
fine grain noise. The wide range of spatial frequencies of noise is independently. Due to the presence of strong correlation among the
treated separately in the approximation and the fine scale using a RGB-colour channels, the solution is far from optimum. The most
multiscale NLM filter. The smoothing parameters, h1 = k1 × σ and plausible solution for such separable algorithms is to convert the
h2 = k2 × σ in (4) is tuned independently of different sub-bands. (correlated) colour representation into a ‘decorrelated’ luminance/
Fig. 4 illustrates the denoised image reconstructed from the colour-difference space like YCbCr or YUV. Literature also
approximation sub-band [coefficients in all other scales are indicates that the ‘decorrelated’ colour-space improves the
assumed to be zero during the reconstruction]. It may be observed denoising performance up to 1.1 dB compared with RGB space
that the coarser grain noises in Fig. 4c are mostly suppressed due to [26]. Moreover, a better subjective quality can be obtained for a
filtering. Similarly, the effect of the NLM filter in suppressing fine- machine vision implementation in YUV colour space compared
grain noise is investigated in Fig. 5. The fine edges and small with RGB space [27]. Therefore, the proposed algorithm, as shown
details of the clean texture image in Fig. 5a are mostly preserved in Fig. 7, is separately applied in YUV space and transformed back

912 IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918


© The Institution of Engineering and Technology 2018
Table 1 Optimised parameter values for (both greyscale and colour) proposed image denoising framework
Parameter Symbols Greyscale parameter Colour algorithm parameter
scale-dependent cont. for threshold k 1.8 1.5 (Y-channel)
1.8 (U & V -channel)
NLM filter smoothing parameter k1 0.2 σ ≤ 10 each YUV channel
0.6 σ > 10 0.2 σ ≤ 10
0.6 σ > 10
k2 0.6 0.6
GIF regularisation parameter k3 3.5 σ ≤ 40 each YUV channel reconstructed RGB image
3.5 σ ≤ 40 0.4 σ ≥ 30
5 σ > 40 5 σ > 40

Fig. 7  Illustration of proposed image denoising framework for colour image

into RGB space to obtain the final denoised image using the colour 4.1 Parameter optimisation
transform, S (inverse transform, S−1, See (15)) matrix as
All the parameters in the proposed image denoising algorithm are
T T uncorrelated to each other. They act independently to improve the
yY, yU, yV = S yR, yG, yB (14) overall performance of the restored image. Therefore, we
separately tuned each parameter at maximum PSNR and SSIM
where yRGB and yYUV is the noisy image in RGB and YUV colour measure for several noise strengths σ on the greyscale images
space, respectively (see (15)) Since the U and V channels are more shown in Fig. 8 (assuming the other parameters are constant). The
corrupted compared with the Y channel, we have selected different average variations in PSNR and SSIM measure for the smoothing
scale dependent constant, k for threshold (as given in Table 1). parameters of NLM filter in the approximation and the fine scale,
However, at higher noise strength, σ ≥ 30 the U and V channels k1 & k2 (Eq.4), the scale-dependent threshold variable k (5) and the
become severely degraded by noise. It requires further processing regularisation parameter of GIF k3 (7) are shown in Figs. 9 & 10
to preserve the finest details of the latent image. Thus, we applied separately for four different noise strengths σ = 10, 25, 50, 75. The
GIF as an additional post-processing step (in the reconstructed optimum value of the scale-dependent constant k as shown in
RGB image) to improve the visual quality of the denoised image at Figs. 9a & 10a ranges between 1.5 and 1.8. Similarly, Figs. 9b, c,
higher noise strength. The same may be ineffective at the lower 10b & c reveal that the NLM filter smoothing parameters must be
noise level. tuned between 0.4 and 0.7, in an approximation sub-band and 0.3–
0.8 in the fine scale, respectively. On the other hand, the
4 Experimental results & discussion regularisation parameter of GIF, k3 in Figs. 9d & 10d provides the
maximum PSNR and SSIM measure at k3 = [2.5, 5.5]. Table 1
The overall performance of the proposed algorithm depends on outlines the final selected values used in our algorithm for both
different regularisation parameters of the filters and the scale- single-channel and colour image denoising.
dependent constant of the curvelet threshold. We investigated the
influence of each parameter to obtain the optimum value for the
highest denoising quality. PSNR and SSIM [28] are adopted in this 4.2 Performance evaluation & discussion
study for quantitative evaluation of image denoising quality. The denoising quality of the proposed algorithm was evaluated
However, for multichannel image denoising, the single channel qualitatively as well as quantitatively for both greyscale and RGB
SSIM measure is extended as the mean of all the indices calculated images. The experiments were conducted on a few standard
for each colour (RGB) channel [29]. In addition, the accuracy of greyscale images as shown in Fig. 8 and on the 24 reference
the proposed algorithm both in terms of denoising quality and run- (colour) images of the TID2008 database [30]. To perpetuate
time complexity is evaluated to measure its suitability in time uniformity in comparison, all the images were resized to 512 × 512
constrained real-time applications. and contaminated with simulated AWGN of standard deviations,
σ = [10, 20, 30, 40, 50, 75]. It may also be noted that the strength of
noise in each RGB channel is considered to be independent and

0.299 0.587 0.114 1 0 1.139


S = −0.147 −0.289 0.436 , S−1 = 1 −0.395 −0.581 . (15)
0.615 −0.515 −0.100 1 2.032 0
IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918 913
© The Institution of Engineering and Technology 2018
Fig. 8  Clean/reference grey images of
(a) Barbara, (b) Boat, (c) Building, (d) Cameraman, (e) Couple, (f) Goldhill, (g) House, (h) Lake, (i) Lena and (j) Peppers

Fig. 10  Plot of average SSIM measure for varying parameter values
Fig. 9  Plot of average PSNR (in dB) measure for varying parameter
(a) Scale-dependant constant for thresholding (k), (b) NLM filter smoothing parameter
values
in approximation sub-band k1 , (c) Fine scale k2 at different noise strengths σ , (d)
(a) Scale-dependant constant for thresholding (k), (b) NLM filter smoothing parameter
in approximation sub-band k1 , (c) Fine scale k2 at different noise strengths σ , (d) GIF regularisation parameter k3 at different noise strengths σ

GIF regularisation parameter k3 at different noise strengths σ


The experimental results on Lena image corrupted with AWGN
equal, i.e. σ  =  σR =  σG =  σB. As illustrated in Section 3.2 the of σ = 40 are shown in Fig. 11 for qualitative assessment. A
portion (patch) of each denoised image is magnified (up to 300%)
proposed separable algorithm is applied independently in YUV
to indicate the efficacy of the proposed technique in suppressing
colour space. Therefore, the covariance matrix of noise was also
noise while producing sharp edges with a minimum loss of fine
computed for YUV colour space
details even at higher noise strength σ = 40 . The quantitative
assessment of the proposed method in terms of mean PSNR and
σR2 0 0
~
SSIM for the set of standard greyscale images (Fig. 8) is given in
T
R = SRS , R= 0 σG2 0 , Table 2. A similar analysis for a colour image in Fig. 12 was also
carried out to exhibit the visual improvement of multichannel
0 0 σB2 image restoration. To complement the qualitative valuation, we
calculated the corresponding PSNR (in dB) and SSIM measures as
where T denotes the transpose of a matrix. shown in Fig. 12. Table 3 illustrates the mean measures on
The denoising quality of the proposed algorithm was evaluated TID2008 database [30]. The following few facts may be noticed
and compared using PSNR (in dB) and SSIM indices. CT [12], from the results
NLM filtering [3], wavelet NeighSURE shrinkage [11], non-sub-
sampled shearlet transformation (NSST)-based hard thresholding • The three-scale decomposition technique for image denoising
[31], patch-based principal components analysis (PL-PCA) [32], distinctively represents the wide spatial frequency
K-means clustering singular value composition (K-SVD) [33], characteristics of noise in the approximation, coarser and fine
MBF [13], NLM method noise thresholding (NLMNT) [17] and scale. The combination of multiscale NLM and curvelet hard
the state-of-the-art BM3D [19] are considered in denoising quality thresholding are distinctively applied in different scales to
comparison for the greyscale image. Similarly, the efficacy of the suppress them independently.
proposed method for colour image is also compared with CT [12], • The ‘edge-aware’ GIF [4] is implemented to preserve the local
NLM [34], MBF [13], multi-channel SURE-LET [26], local pixel structures like edges, textures and small details in the restored
grouping (LPG)-PCA [29] and colour-block matching, sparse 3D image, as shown in Fig. 6.
collaborative filtering (CBM3D) [35].

914 IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918


© The Institution of Engineering and Technology 2018
Fig. 11  Denoised image obtained using different algorithms for Lena image corrupted with AWGN of σ = 40. The quantitative measures for respective
denoised images are
(a) CT [12]: PSNR = 28.6551, SSIM = 0.832, (b) NLM [3]: PSNR = 28.129, SSIM = 0.785, (c) NeighSURE [11]: PSNR = 29.4010, SSIM = 0.836, (d) NSST [31]: PSNR = 30.390,
SSIM = 0.867, (e) PL-PCA [32]: PSNR = 30.021, SSIM = 0.868, (f) K-SVD [33]: PSNR = 29.746, SSIM = 0.849, (g) MBF [13]: PSNR = 28.849, SSIM = 0.833, (h) HLMNT [17]:
PSNR = 29.0802, SSIM = 0.807, (i) BM3D [20]: PSNR = 29.612, SSIM = 0.854 and (j) Proposed method: PSNR = 30.237, SSIM = 0.875. A patch of each image is magnified for
visual assessment

Fig. 12  Quantitative assessments for the respective denoised images are
(a) Corrupted image σ = 40 : PSNR: 16.093, SSIM: 0.359, (b) CT [12]: PSNR: 29.116, SSIM: 0.830, (c) NLM [34]: PSNR: 27.507, SSIM: 0.758, (d) MBF [13]: PSNR: 26.413,
SSIM: 0.816, (e) MC-SURE-LET [26]: PSNR: 30.498, SSIM: 0.846, (f) LPG-PCA [29]: PSNR: 30.686, SSIM: 0.819, (g) CBM3D [35]: PSNR: 31.113, SSIM: 0.865, (h) Proposed
method: PSNR: 31.135, SSIM: 0.870. A patch of each image is magnified for better visual interpretation

• The spatial edge-preserving filtering in curvelet domain suppressing visible colour artefacts near smooth regions while
combined with thresholding enhances the performance of image preserving the sharp image edges and texture details. Moreover, it
denoising at all noise strengths σ compared with the individual is also known that the denoising quality of GIF depends on another
methods of NLM filter [3] and CT [12]. external input image called the guidance image. Here, the denoised
• Compared with other combined approaches in [13, 17], the image in Fig. 13c is restored by utilising the local features of the
proposed technique improves the denoising performance at a flash image in Fig. 13a (which acts as the guidance image, G).
higher noise strengths and provides better and comparable However, the proposed method denoise the image by considering
results – with BM3D [19], GSRC [21] & CBM3D [35] – at the content of the input noisy image, exclusively.
lower σ for both grey and colour image. The anisotropic scaling of curvelet to represent edges (C2
smooth) with optimal sparsity and the high intra-scale similarity
For completeness, the competitiveness of the proposed among the coefficients motivated the authors to consider NLM as a
algorithm is compared with the GIF [4], as shown in Fig. 13. The multiscale filter to suppress noise in different scales. Unlike CT
results indicate the superiority of the proposed method in [12], the proposed multiscale filtering is able to distinguish

IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918 915


© The Institution of Engineering and Technology 2018
Fig. 13  Image denoising performance comparison between the GIF and the proposed approach
(a) Noise-free guidance image (with flash), (b) Input/no-flash image (PSNR = 13.852, SSIM = 0.2254), (c) Denoised image using GIF (PSNR = 14.719, SSIM = 0.435), (d) Denoised
image using proposed approach (PSNR = 14.821, SSIM = 0.457). The quantitative measures are obtained by assuming the flash image as the reference image

Table 2 Mean PSNR (dB) and SSIM measure between original and restored images for different denoising techniques on test
image in Fig. 8
PSNR, dB SSIM [28]
Denoised methods σ = 10 σ = 20 σ = 30 σ = 40 σ = 50 σ = 75 σ = 10 σ = 20 σ = 30 σ = 40 σ = 50 σ = 75
CT [12] 32.592 29.594 28.043 26.960 26.134 24.631 0.938 0.880 0.834 0.797 0.766 0.706
NLM [3] 35.103 31.452 29.035 27.252 25.836 23.167 0.956 0.901 0.839 0.776 0.716 0.585
NeighSURE [11] 34.732 31.022 28.978 27.621 26.609 24.910 0.956 0.905 0.858 0.816 0.780 0.707
NSST [31] 35.284 32.027 30.173 28.885 27.874 26.134 0.959 0.919 0.884 0.852 0.822 0.758
PL-PCA [32] 35.608 32.238 30.026 29.069 27.691 26.002 0.964 0.920 0.887 0.857 0.822 0.754
K-SVD [33] 36.129 32.641 30.484 28.930 27.684 25.487 0.963 0.921 0.879 0.838 0.799 0.711
MBF [13] 33.386 30.107 28.322 27.142 26.291 24.875 0.934 0.882 0.841 0.805 0.773 0.700
NLMNT [17] 34.744 31.690 29.597 27.917 26.501 23.699 0.958 0.907 0.853 0.798 0.746 0.635
BM3D [19] 36.152 32.703 30.574 28.766 27.868 25.629 0.965 0.928 0.889 0.850 0.819 0.737
GSRC [21] 35.373 32.324 30.652 29.024 28.153 26.129 0.958 0.918 0.887 0.858 0.823 0.752
Proposed 35.443 32.081 30.526 29.089 27.927 26.261 0.962 0.919 0.896 0.864 0.826 0.759

Table 3 Mean PSNR (dB) and SSIM measure between original and restored images for different denoising techniques on
TID2008 database [30]
PSNR, dB SSIM [28, 29]
Denoising methods σ = 10 σ = 20 σ = 30 σ = 40 σ = 50 σ = 75 σ = 10 σ = 20 σ = 30 σ = 40 σ = 50 σ = 75
CT [12] 32.118 29.117 27.523 26.474 25.698 24.322 0.883 0.809 0.755 0.715 0.684 0.629
NLM [34] 33.577 30.089 27.864 26.151 24.743 22.015 0.9044 0.821 0.741 0.691 0.603 0.566
MBF [13] 32.634 29.050 26.763 24.973 23.459 20.466 0.883 0.801 0.741 0.694 0.655 0.583
SURE-LET [26] 35.723 31.897 29.820 28.429 27.402 25.654 0.935 0.870 0.817 0.772 0.735 0.662
LPG-PCA [29] 34.974 30.894 28.742 27.279 26.215 24.287 0.935 0.862 0.799 0.743 0.694 0.589
CBM3D [35] 36.546 32.719 30.558 29.046 28.052 26.048 0.943 0.882 0.822 0.766 0.733 0.627
Proposed 35.682 32.142 30.154 28.953 28.143 26.204 0.9318 0.874 0.829 0.777 0.745 0.673

between fine edges and noises both in the approximation and the The run-time complexity is compared with the CT technique
fine scale. On the other hand, the spatial domain NLM filter search [12], BM3D [19], MBF [13] and GSRC [21] techniques. The
for similar patches in a larger neighbouring window (which is experiment was conducted on a greyscale image of dimensions
again local [23]) and the similarity among patches further 1024 × 1024, 512 × 512, 256 × 256 and 128 × 128, respectively.
decreases with higher noise strength. Similarly, the inability of BF Each algorithm was executed at least 50 times for four different
in preserving gradient direction and the limitations of the wavelet noise levels, σ = 10, 25, 50, 75. The runtime was computed on a
in representing image edges degrade the performance of MBF [13]. 4.00 GB RAM, 64-bit Intel(R) Core™, i7, 4770 CPU @ 3.40 GHz
The patch-based methods like PL-PCA, LPG-PCA, BM3D and PC using MATLAB R2015a. Fig. 14 illustrates the mean–standard
recently proposed GSRC search for more number of patches with deviation plot of execution time (in seconds) for the corrupted
similar local spatial structures that decrease at a higher noise level Lena image denoised using the above methods. It can be observed
and thus restricts the performance of denoising. However, the that the use of fast filtering algorithms and the limited
competitiveness of the proposed algorithm is comparable with decomposition level of the proposed method accelerates the
several state-of-the-art techniques at lower noise strength but denoising process and it is comparable with [12]. Unlike, BM3D
improves at a higher value of σ for both grey and colour images. [19] and GSRC [21], the complexity of the proposed technique is
independent of noise standard deviation. Moreover, the iterative
4.3 Computational complexity GSRC technique achieves a modest increase in denoising quality
(at higher noise strength) with higher implementation complexity.
The use of fast NLM [34] and GIF O(N) [4] algorithms in our
CT-based hybrid approach aides in minimising the overall
complexity of image denoising. Moreover, the speed of the
5 Conclusion
operation of the proposed method is accelerated by limiting the Most image denoising methods assume that the signal is smooth or
number of decomposition levels to Nγ = 3. piecewise smooth and the noise is oscillatory. However, many fine
structures and small details of an image are as oscillatory as noise,

916 IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918


© The Institution of Engineering and Technology 2018
Fig. 14  Run-time (in s) of MBF [13] (red), BM3D [19] (blue), CT [12] (green), proposed algorithm (magenta) and GSRC [21] (black) for image dimension
of
(a) 1024 × 1024, (b) 512 × 512, (c) 256 × 256 and (d) 128 × 128, respectively, are computed for the AWGN of std. σ = 10, 25, 50, 75

which poses challenges in image modelling. The proposed signal [12] Starck, J.L., Candès, E.J., Donoho, D.L.: ‘The curvelet transform for image
denoising’, IEEE Trans. Image Process., 2002, 11, (6), pp. 670–684
modelling analysed the image in three curvelet scales to denoise [13] Zhang, M., Gunturk, B.K.: ‘Multiresolution bilateral filtering for image
both the smooth (low-frequency) and oscillatory (high-frequency) denoising’, IEEE Trans. Image Process., 2008, 17, (12), pp. 2324–2333
noise. Unlike spatial domain implementation, curvelet domain [14] Knaus, C., Zwicker, M.: ‘Dual-domain image denoising’. 2013 20th IEEE Int.
implementation of multiscale NLM filter extracts more adequately Conf. on Image Processing (ICIP), 2013, pp. 440–444
[15] Ma, J., Plonka, G.: ‘The curvelet transform’, IEEE Signal Process. Mag.,
the self-similarity among the high correlated coefficients within 2010, 27, (2), pp. 118–133
individual scales. The visual and quantitative results yield the [16] Wu, K., Zhang, X., Ding, M.: ‘Curvelet based non-local means algorithm for
potential of multiscale NLM in exploiting self-similarity among the image denoising’, AEU-Int. J. Electron. Commun., 2014, 68, (1), pp. 37–43
patches (in curvelet domain) even at higher noise strength [17] Kumar, B.S.: ‘Image denoising based on non-local means filter and its method
noise thresholding’, Signal Image Video Process., 2013, 7, (6), pp. 1211–1227
σ ≥ 40 . It is also observed that the implementation of edge [18] Coupé, P., Manjón, J.V., Robles, M., et al.: ‘Adaptive multiresolution non-
preserving (NLM) filter – instead of hard thresholding – in the local means filter for three-dimensional magnetic resonance image
finer scale aides in retaining well-connected edges with small denoising’, IET Image Process., 2012, 6, (5), pp. 558–568
image details. On the other hand, the application of hard [19] Dabov, K., Foi, A., Katkovnik, V., et al.: ‘Image denoising by sparse 3-D
transform-domain collaborative filtering’, IEEE Trans. Image Process., 2007,
thresholding in the coarser scale removed the insignificant noise 16, (8), pp. 2080–2095
coefficients from the high magnitude signals. The inevitable [20] Knaus, C., Zwicker, M.: ‘Progressive image denoising’, IEEE Trans. Image
ringing artefacts in the reconstructed image (which is mostly Process., 2014, 23, (7), pp. 3114–3125
ignored in many transform domain techniques) is further processed [21] Zha, Z., Liu, X., Zhou, Z., et al.: ‘Image denoising via group sparsity residual
constraint’ IEEE Int. Conf. on Acoustics, Speech and Signal Processing
using GIF to keep the sharp image edges with a minimum loss of (ICASSP), New Orleans, LA, 2017, pp. 1787–1791, doi: 10.1109/
fine details. The results on both greyscale and colour images have ICASSP.2017.7952464
shown encouraging quantitative and qualitative improvement [22] Karami, A., Tafakori, L.: ‘Image denoising using generalised Cauchy filter’,
compared with several state-of-the-art techniques at higher noise IET Image Process., 2017, 11, (9), pp. 767–776
[23] Buades, A., Coll, B., Morel, J.M.: ‘Non-local means denoising’, Image
strength. However, the proposed technique unable to preserve the Process. Online, 2011, 1, pp. 208–212
fine-scale details at lower noise strength requires further study and [24] Deledalle, C.A., Duval, V., Salmon, S.: ‘Non-local methods with shape-
it is left as the future scope of the work. adaptive patches (NLM-SAP)’, J. Math. Imag. Vis., 2012, 43, (2), pp. 103–
120
[25] Alecu, A., Munteanu, A., Pizurica, A., et al.: ‘Information-theoretic analysis
6 References of dependencies between curvelet coefficients’. IEEE Int. Conf. on Image
Processing, 2006, pp. 1617–1620
[1] Milanfar, P.: ‘A tour of modern image filtering: new insights and methods, [26] Luisier, F., Blu, T.: ‘SURE-LET multichannel image denoising: interscale
both practical and theoretical’, IEEE Signal Process. Mag., 2013, 30, (1), pp. orthonormal wavelet thresholding’, IEEE Trans. Image Process., 2008, 17,
106–128 (4), pp. 482–492
[2] Tomasi, C., Manduchi, R.: ‘Bilateral filtering for gray and color images’. [27] Podpora, M., Korbas, G.P., Kawala-Janik, A.: ‘YUV vs RGB-choosing a
Sixth Int. Conf. on Computer Vision, 1998, pp. 839–846 color space for human-machine interaction’. Position papers of the 2014
[3] Buades, A., Coll, B., Morel, J.M.: ‘A review of image denoising algorithms, Federated Conf. on Computer Science and Information Systems, 2014, pp.
with a new one’, Multiscale Model. Simul., 2005, 4, (2), pp. 490–530 29–34
[4] He, K., Sun, J., Tang, X..: ‘Guided image filtering’, IEEE Trans. Pattern [28] Wang, Z., Bovik, A.C., Sheikh, H.R., et al.: ‘Image quality assessment: from
Anal. Mach. Intell., 2013, 35, (6), pp. 1397–1409 error visibility to structural similarity’, IEEE Trans. Image Process., 2004, 13,
[5] Donoho, D.L., Johnstone, J.M.: ‘Ideal spatial adaptation by wavelet (4), pp. 600–612
shrinkage’, Biometrika, 1994, 81, (3), pp. 425–455 [29] Zhang, L., Dong, W., Zhang, D., et al.: ‘Two-stage image denoising by
[6] Donoho, D.L., Johnstone, I.M., Kerkyacharian, G., et al.: ‘Wavelet shrinkage: principal component analysis with local pixel grouping’, Pattern Recognit.,
asymptopia?’, J. R. Stat. Soc. B, Methodol., 1995, 57, (2), pp. 301–369 2010, 43, (4), pp. 1531–1549
[7] Chang, S.G., Yu, B., Vetterli, M.: ‘Adaptive wavelet thresholding for image [30] Ponomarenko, N., Lukin, V., Zelensky, A., et al.: ‘Tid2008-a database for
denoising and compression’, IEEE Trans. Image Process., 2000, 9, (9), pp. evaluation of full-reference visual quality assessment metrics’, Adv. Mod.
1532–1546 Radioelectron., 2009, 10, (4), pp. 30–45
[8] Sendur, L., Selesnick, I.W.: ‘Bivariate shrinkage with local variance [31] Easley, G., Labate, D., Lim, W.Q.: ‘Sparse directional image representations
estimation’, IEEE Signal Process. Lett., 2002, 9, (12), pp. 438–441 using the discrete shearlet transform’, Appl. Comput. Harmon. Anal., 2008,
[9] Pizurica, A., Philips, W.: ‘Estimating the probability of the presence of a 21, (1), pp. 25–46
signal of interest in multiresolution single- and multiband image denoising’, [32] Deledalle, C.A., Salmon, J., Dalalyan, A.S.: ‘Image denoising with patch
IEEE Trans. Image Process., 2006, 15, (3), pp. 654–665 based PCA: local versus global’. British Machine Vision Conf. (BMVC),
[10] Blu, T., Luisier, F.: ‘The SURE-LET approach to image denoising’, IEEE 2011, vol. 81, pp. 425–455
Trans. Image Process., 2007, 16, (11), pp. 2778–2786
[11] Dengwen, Z., Wengang, C.: ‘Image denoising with an optimal threshold and
neighbouring window’, Pattern Recognit. Lett., 2008, 29, (11), pp. 1694–1697

IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918 917


© The Institution of Engineering and Technology 2018
[33] Easley, G., Labate, D., Lim, W.Q.: ‘k -SVD: An algorithm for designing [35] Dabov, K., Foi, A., Katkovnik, V., et al.: ‘Color image denoising via sparse
overcomplete dictionaries for sparse representation’, IEEE Trans. Signal 3D collaborative filtering with grouping constraint in luminance-chrominance
Process., 2006, 54, (11), pp. 4311–4322 space’. IEEE Int. Conf. on Image Processing (ICIP), 2007, vol. 1, pp. I–313
[34] Kroon, D.J.: ‘Fast Non-local means 1D, 2D color and 3D’, https://
in.mathworks.com/matlabcentral/fileexchange/27395-fast-non-local-
means-1d--2d-color-and-3d, accessed April 2017

918 IET Image Process., 2018, Vol. 12 Iss. 6, pp. 909-918


© The Institution of Engineering and Technology 2018

También podría gustarte