Available online at www.sciencedirect.

com

ScienceDirect
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2017) 000–000
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 115 (2017) 526–532

7th International Conference on Advances in Computing & Communications, ICACC-2017, 22-

24 August 2017, Cochin, India

Constrained least squares filtering followed by denoising of

decomposed images using wave atom and wavelet transform
S Shruthi*, K G Satheeshkumar
Amal Jyothi College of Engineering, Koovappally P.O, Kanjirappally, 686518, India

Abstract

Image restoration is very essential in order to recover an image from its blurred and noisy form. A novel algorithm is proposed
here which uses constrained least squares filter for deconvolution followed by denoising section. Deblurred image obtained after
deconvolution is decomposed into texture and cartoon parts, using curvelet transform, which are then denoised using wave atom
transform and wavelet transform respectively. Cycle spinning technique is used to enhance the denoising capability of wavelets
and wave atoms. This algorithm does not require a priori knowledge of the noise distribution and provides effective noise reduction
for texture dominant images.

© 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing &
Communications.

Keywords: Constrained least squares filtering; curvelet transform; cycle spinning; image decomposition; wave atom transform; wavelet transform.

1. Introduction

Image restoration is the process of recovering an image from its degraded form. Different techniques such as
deblurring and denoising are used for this purpose. Deconvolution methods minimize blur as well as noise. If the
information regarding the point spread function is not known, deconvolution is said to be blind and if it is known, then
it is said to be non-blind.

* Corresponding author. Tel.: 09400646390.

E-mail address: shruthiskedar@gmail.com

1877-0509 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing &
Communications.

1877-0509 © 2017 The Authors. Published by Elsevier B.V.

Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing &
Communications
10.1016/j.procs.2017.09.110
 S Shruthi et al. / Procedia Computer Science 115 (2017) 526–532 527
2 S Shruthi et al. / Procedia Computer Science 00 (2017) 000–000

Parametric Wiener filtering followed by wave atom transform [1], denoising by supervised adaptive fusion of
decomposed images [2] are some of the recent related works. Image denoising based on wave atoms and cycle spinning
[3] shows the advantage of using cycle spinning. In another work, variable splitting fast image recovery (VSFIR) has
been introduced to obtain an equivalent constrained optimization formulation, addressed with an augmented
Lagrangian method [4]. TV-BDSB is a new TV (total variation) blind deconvolution method which is employed using
split Bregman iteration [5]. For non-local TV regularization problems, algorithm is presented based on operator
splitting technique and Bregman iteration [6]. Non-blind image deblurring (NBID) is a method that has been
introduced to preserve the salient edges during restoration [7].
Restoration using Wiener filtering [8], [9] requires both the information regarding noise as well as that of blur to
obtain the estimate of original image. The proposed work deals with the restoration of degraded images where the
information regarding the noise is unknown. Constrained least squares filter is hence used here for deconvolution.
After filtering process, the deblurred image is decomposed into texture and cartoon parts using curvelet transform. The
decomposed texture part is denoised using wave atom transform taking the advantage of cycle spinning while the
decomposed cartoon part is denoised using wavelet transform with cycle spinning. The denoised texture and cartoon
parts are then fused to form the final denoised image. Cycle spinning is a technique that has been used in order to
overcome the distortions that occur during denoising by coefficient thresholding. Wave atoms denoise oscillatory
patterns and textures while wavelets denoise cartoon parts in an image effectively. So combining constrained least
squares filter, wave atoms, and wavelets along with cycle spinning provides better restoration results.

2. Description

2.1. Degradation and Wiener filter

The process in which an image is affected by blur and noise is called degradation. Let f be the clear image degraded
by a degradation function H, and additive noise n, thus a noisy blurred image g is formed.

g  f *H n (1)

Wiener filter requires statistical noise characteristics and degradation function for image restoration process where
noise and image are assumed to be uncorrelated. The objective of this filter is to find the estimate of clear image f to
minimize the mean square error. Frequency domain solution for Wiener filter is given by the following transfer
function:

 
 H *
(u , v ) 
Fˆ (u, v)   G (u, v) (2)
 H (u, v) 2  N (u, v) 
 I (u, v) 

H (u, v) is the degradation function, H* (u, v) is its complex conjugate, G (u, v) is the noisy blurred image, N (u, v)
is the noise power spectrum and I (u, v) is the power spectrum of original image. N (u, v) / I (u, v) is known as noise
to signal power ratio.

2.2. Constrained least squares filter

Constrained least squares filter is an extension of Wiener filter where the deconvolution does not require
information of the noise. Its frequency domain solution is given by the following transfer function:
528 S Shruthi et al. / Procedia Computer Science 115 (2017) 526–532
S Shruthi et al. / Procedia Computer Science 00 (2017) 000–000 3

 H * (u, v) 
Fˆ (u, v)   G (u, v) (3)
2 2
 H (u, v)   P(u , v) 

is a parameter that has to be manually adjusted to get best visual results and P (u, v) is the Fourier transform of the
Laplacian operator p (x, y):

 0 1 0 
p ( x, y )   1 4  1 (4)
 0  1 0 

Constrained least squares filter provides better results [8] while comparing with Wiener filter for high and medium
noise, and for low noise, results are almost equal.

2.3. Wavelet transform

Wavelet transform [9], [10], which provides time frequency representation, was introduced to overcome the
limitations of short time Fourier transform (STFT). STFT gives constant resolution for all frequencies while wavelet
uses a technique called multi resolution in which different resolutions are used to analyze different frequencies. For
spatially localized details like singularities and edges, wavelet transform attain good sparsity because such details
provide a significant portion of the information and are abundant in natural images. Therefore, it has found a major
application for denoising of images. Wavelet transform is classified as continuous wavelet transform (CWT) and
discrete wavelet transform (DWT). DWT can be developed using orthogonal wavelets and non-orthogonal wavelets
can be used for developing CWT. Multilevel DWT using cycle spinning is implemented here for denoising.

2.4. Curvelet transform

Curvelet transform has been developed in order to overcome the shortcomings of wavelets. It is a multiscale
pyramid and at each length scale have several directions and positions, and at fine scales have needle-shaped elements.
Curvelet transform is of continuous and discrete type. The Cartesian coronae of continuous curvelet transform is based
on concentric circles while that of discrete curvelet transform is based on concentric squares and shears. Wrapping
and USFFT (Unequispaced FFT) are novel methods of fast discrete curvelet transforms (FDCTs) [11]. These
transforms are simpler, less redundant and faster than the existing curvelet methods. They differ in the way of
translation of curvelets. For USFFT, translation grid is tilted so that it is aligned with the curvelet orientation and in
the case of wrapping, for all angle within each quadrant grid is the same. Wrapping transform is simpler to implement
and understand. FDCT via wrapping is implemented here for decomposition.

2.5. Wave atom transform

Wave atom transform is an effective denoising technique for images containing oscillatory patterns and textures.
It performs better than wavelets and curvelets [11] for such images. It obeys parabolic scaling relation [12] i.e.
wavelength ~ (diameter)2, interpolating accurately between directional wavelets and Gabor. Wave atoms are capable
of adjusting to arbitrary local directions of patterns, and anisotropic patterns which are aligned with the axes are
sparsely represented by them.
The geometry of the tiling of wave packets must be approximately polar in frequency and approximately Cartesian
in space for a tiling to be universal and the universality suggests two parameters α and β for indexing wave packet
architecture.  specifies whether the decomposition is multiscale (=1) or not (=0) while  indicate whether the
basis are localized and poorly directional (=1) or extended and fully directional (=0). =1 and =1/2 for curvelets,
 S Shruthi et al. / Procedia Computer Science 115 (2017) 526–532 529
4 S Shruthi et al. / Procedia Computer Science 00 (2017) 000–000

wavelets have ==1, Gabor transform corresponds to ==0 and for ridgelets =1,  =0. In case of wave atoms,
= =1/2. Fig. 1 shows the relation between the parameters andfor different transforms of wave packet families.

Fig. 1: Various transforms of wave packet families [12]

2.6. Cycle spinning

Cycle spinning [3], [13] is a technique used to avoid distortions that occur during denoising. In the threshold
denoising process, pseudo-Gibbs phenomenon will be produced in the image edges and textures due to lack of
translation invariance of wave atoms. This phenomenon is therefore controlled efficiently when the concept of cycle
spinning is used for wave atom denoising, and thus can get better visual effect as well as PSNR (Peak signal to noise
ratio) compared with wave atoms alone [3]. Wavelet denoising also provides better results using cycle spinning.

2.7. Proposed System

The proposed system consists of degradation and restoration sections as shown in Fig. 2. Degradation is the process
in which blur and noise are added to a clear image and restoration is the recovery process. This system combines
constrained least squares filtering and denoising of decomposed images using wave atom transform and wavelet
transform along with cycle spinning. The original image is first degraded by convolving it with motion blur of specific
length and angle parameter, followed by addition of zero mean white Gaussian noise. Motion blur is the degradation
that occurs in an image due to system movement or shake. The length and angle parameters are used to generate the
point spread function or blurring kernel. Zero mean Gaussian noise of a particular BSNR is then added to it. The
blurred noisy image thus obtained is first deconvoluted by constrained least squares filtering in Fourier space. The
deblurred image thus obtained is then denoised.
To perform denoising, the filtered image is decomposed into texture and cartoon parts using curvelet transform.
For curvelet transform, different scales represent different frequencies that are different filters. Retaining coefficients
corresponding to higher scales and setting lower scale coefficients to zero provides noisy texture image and subtracting
this image from filtered image gives noisy cartoon image. Wave atom transform is applied to denoise noisy texture
part and wavelet transform is applied to denoise noisy cartoon part. Cycle spinning [13] is used for wave atom
denoising and wavelet denoising to attain better result. Circular shift [12], [13] is used for implementing cycle
spinning. Hard thresholding (Universal Threshold) is used in this work because too much edge details are lost if soft
thresholding is used.

T   2 log N (5)
530 S Shruthi et al. / Procedia Computer Science 115 (2017) 526–532
S Shruthi et al. / Procedia Computer Science 00 (2017) 000–000 5

T is the hard threshold, N is the image size, is the noise variance. Noise variance is calculated by using normalized
histogram.

Fig. 2. Block diagram of proposed system

L 1 2

   ( f i  m) P ( f i ) (6)
i 0

fi denotes intensity values in an image, P (fi) represents its corresponding probability of occurrence, m is the mean of
the image and L is the total number of intensity values.
After thresholding, inverse wave atom transform and inverse wavelet transform are determined using reverse cycle
spinning. The denoised texture components and denoised cartoon components are added to obtain restored image.
This method provides better results without the knowledge regarding the noise having higher SSIM (Structural
similarity index measure) and PSNR values.

2.8. Experimental Results

Images of 256 x 256 size are degraded using motion blur with a particular length parameter and angle parameter.
Here the length parameter=15 and theta=135 are taken for comparison. To this blurred image, zero mean Gaussian
noise having BSNR=20 dB is added to it. This degraded image undergoes constrained least squares filtering followed
by the decomposition using curvelet transform and denoising using wavelet, and wave atom transform with cycle
spinning. SSIM (Structural Similarity Index Measure) and PSNR (Peak Signal to Noise Ratio) are used here for
comparison of results. PSNR is the noise suppression ability. Higher PSNR value implies better noise suppression
which is calculated using mean square error (MSE).

 M2  (7)
PSNR  10 log10  
 MSE 

M is the maximum pixel value of the image. SSIM is the quality of visual perception of the image and its maximum
value is 1. Denoising is said to be better if SSIM is closer to 1.
6 S Shruthi
S Shruthi et al. / Procedia
et al. / Procedia Computer
Computer ScienceScience 115000–000
00 (2017) (2017) 526–532 531

(2 x  y  c1 )(2 xy  c 2 ) (8)

SSIM ( x, y ) 
(  x2   y2  c1 )( x2   y2  c 2 )

x refers to the average of x, y is the average of y, x refers to the variance of x, y refers to the variance of y,xy is
the covariance of x and y. Two variables c1 and c2 where c1 = (K1L) 2 and c2 = (K2L) 2 are used to stabilize the division.
K1=0.01 and K2=0.03 and L is the dynamic range of pixel values. L=255 for 8 bit image.
Experiments are performed using Matlab R2013a with the help of CurveLab-2.1.3 and WaveAtom-1.1.1 toolbox.
Table 1 gives the comparison of PSNR values (dB) and Table 2 gives the comparison of SSIM values of proposed
work with previous works for Cameraman, Lena and Barbara images. From the values obtained, it can be seen that
texture dominant images like Lena and Barbara images gives higher values of PSNR and SSIM while Cameraman
image, which is not texture dominant, does not give highest value for PSNR and SSIM. The same algorithm is used
for restoring an image taken from a mall as shown in Fig. 3. Restoration of noisy blurred image in Fig. 3 gives
PSNR=32.80 dB and SSIM=0.906.

Table 1. Comparison of PSNR in dB.

Images NBID method Parametric Wiener Proposed method
(dB) filter and wave (dB)
atom transform
(dB)
Cameraman 26.88 27.96 27.11
Lena 27.73 27.95 31.59
Barbara 22.90 23.51 28.85

Table 2. Comparison of SSIM.

Images NBID method Parametric Wiener Proposed method
filter and wave
atom transform
Cameraman 0.815 0.636 0.813
Lena 0.838 0.751 0.906
Barbara 0.711 0.720 0.886

Fig. 3. (a) Original image; (b) Noisy blurred image; (c) Restored image.

These results show that the algorithm restores all types of images and gives a better performance for images which
are texture dominant.
532 S Shruthi et al. / Procedia Computer Science 115 (2017) 526–532
S Shruthi et al. / Procedia Computer Science 00 (2017) 000–000 7

3. Conclusion

Efficient image restoration is obtained from the proposed algorithm in which constrained least squares filtering
followed by denoising of decomposed images using wave atom transform and wavelet transform combining cycle
spinning results in better PSNR and SSIM for texture dominant images.

References

[1] Z. Mbarki, H. Seddik, E. B. Braiek. A rapid hybrid algorithm for image restoration combining parametric Wiener filtering and wave atom
transform, J. Vis. Commun. Image R. 2016; 40: 694-707.
[2] P. D. Swami, A. Jain. Image denoising by supervised adaptive fusion of decomposed images restored using wave atom, curvelet and wavelet
transform, SIViP 2014; 8 (3): 443-459.
[3] W. Q. Zhang, Y. M. Song, J. Q. Feng. A new image denoising method based on wave atoms and cycle spinning, JSW 2014; 9: 216-221.
[4] M.V. Afonso, J.M. Bioucas-Dias, M.A.T. Figueiredo. Fast image recovery using variable splitting and constrained optimization, IEEE
Transactions on Image Processing 2010; 19 (9): 2345-2356.
[5] W.H. Li, Q.L. Li, W.G. Gong, S. Tang. Total variation blind deconvolution employing split Bregman iteration, J. Vis. Commun. Image Rep.
2012; 23 (3): 409-417.
[6] X.Q. Zhang, M. Burger, X. Bresson, S. Osher. Bregmanized non local regularization for deconvolution and sparse reconstruction, SIAM J.
Imag. Sci. 2010; 3(3): 253-276.
[7] S. Tang, W. Gong, W. Li, W. Wang. Non-blind image deblurring method by local and nonlocal total variation models, Signal Process. 2014;
94: 339-349.
[8] R. C. Gonzalez, R. E. Woods. Digital Image Processing; New Jersey, Prentice Hall; 2002.
[9] S Jayaraman, S Esakkirajan, T Veerakumar. Digital Image Processing; New Delhi, McGraw Hill Education Private Limited; 2013.
[10] K. Ding. Wavelets, curvelets and wave atoms for image denoising. 3rd International Congress on Image and Signal Processing 2010; 2: 782-
786.
[11] E. Candes, L. Demanet, D. Donoho, and L. Ying. Fast discrete curvelet transforms, SIAM Multiscale Model. Simul. 2006; 5 (3): 861-899.
[12] L. Demanet, L. Ying. Wave atoms and sparsity of oscillatory patterns, Applied and Computational Harmonic Analysis 2007; 23 (3): 368-387.
[13] M. Misiti, Y. Misiti, G. Oppenheim, J. M. Poggi. Wavelet toolbox user's guide, September 2015.