Multiplicative noise removal with a sparsity-aware optimization model

Jian Lu; Lixin Shen; Chen Xu; Yuesheng Xu

doi:10.3934/ipi.2017044

Article Contents

2017, Volume 11, Issue 6: 949-974. Doi: 10.3934/ipi.2017044

This issue Previous Article Wavelet tight frame and prior image-based image reconstruction from limited-angle projection data Next Article A numerical study of a mean curvature denoising model using a novel augmented Lagrangian method

Multiplicative noise removal with a sparsity-aware optimization model

1.
College of Mathematics and Statistics, Shenzhen University, Shenzhen 518060, China
2.
Department of Mathematics, Syracuse University, Syracuse, NY 13244, USA
3.
School of Data and Computer Science, Guangdong Provincial Key Lab of Computational Science, Sun Yat-sen University, Guangzhou 510275, China
4.
Department of Mathematics and Statistics, Old Dominion University, Norfolk, VA 23529, USA

* Corresponding author: Lixin Shen
* Corresponding author: Lixin Shen

Received: March 31, 2016

Revised: July 31, 2017

Published: November 2017

Abstract Full Text(HTML) Figure(6) / Table(4) Related Papers Cited by

Abstract

Restoration of images contaminated by multiplicative noise (also known as speckle noise) is a key issue in coherent image processing. Notice that images under consideration are often highly compressible in certain suitably chosen transform domains. By exploring this intrinsic feature embedded in images, this paper introduces a variational restoration model for multiplicative noise reduction that consists of a term reflecting the observed image and multiplicative noise, a quadratic term measuring the closeness of the underlying image in a transform domain to a sparse vector, and a sparse regularizer for removing multiplicative noise. Being different from popular existing models which focus on pursuing convexity, the proposed sparsity-aware model may be nonconvex depending on the conditions of the parameters of the model for achieving the optimal denoising performance. An algorithm for finding a critical point of the objective function of the model is developed based on coupled fixed-point equations expressed in terms of the proximity operator of functions that appear in the objective function. Convergence analysis of the algorithm is provided. Experimental results are shown to demonstrate that the proposed iterative algorithm is sensitive to some initializations for obtaining the best restoration results. We observe that the proposed method with SAR-BM3D filtering images as initial estimates can remarkably outperform several state-of-art methods in terms of the quality of the restored images.

Keywords:

Mathematics Subject Classification: Primary: 94A08; Secondary: 49N45, 68U10.

Citation:

Full Text(HTML)

Figure 1. Five gray-level test images. (a) "Cameraman" ($512\times 512$). (b) "Lena" ($512\times 512$). (c) "Peppers" ($512\times 512$). (d) "Remote1" ($768\times 574$). (e) "Remote2" ($632\times 540$)

Download: Full-size image PowerPoint slide

Figure 2. (a) PSNR versus number of iterations. (b) Relative error versus number of iterations. Here, various $x^{(0)}$s are used as initial estimates in Algorithm 1. From top to bottom: the test images are the degraded "Cameraman" with multiplicative noise at levels $L=10$, $6$, $4$, and $2$

Download: Full-size image PowerPoint slide

Figure 3. (a) PSNR versus number of iterations for $L=4$. (b) PSNR versus number of iterations for $L=2$. Here, various $x^{(0)}$s are used as initializations in Algorithm 1 for the degraded "Cameraman" with multiplicative noise

Download: Full-size image PowerPoint slide

Figure 4. (a) PSNR versus number of iterations. (b) Relative error versus number of iterations. Here, two different $x^{(0)}$s are used as initial estimates in Algorithm 1. The solid lines are plotted firstly by selecting the parameters of Algorithm 1 to obtain (nearly) optimal PSNR values (marked by '$\circ$') under the prescribed tolerance $\text{TOL}=3\times 10^{-4}$; then the dashed lines are plotted by using the same parameters as those of corresponding solid lines. The test images are the degraded "Cameraman" with multiplicative noise at various noise levels ($L=10$, $6$, $4$, and $2$)

Download: Full-size image PowerPoint slide

Figure 5. Results of various denoising methods on "Cameraman" image corrupted by multiplicative noise with $L=2$ (the first column) and $L=4$ (the second column). From top to bottom: Noisy images (8.63 dB, 11.64 dB), DZ (24.27 dB, 25.85 dB), TwL-4V (25.28 dB, 26.72dB), Ⅰ-DIV (24.98 dB, 26.52 dB), HMNZ (25.30 dB, 27.33 dB), and Ours (25.47 dB, 27.38 dB)

Download: Full-size image PowerPoint slide

Figure 6. Results of various denoising methods on "Remote1" image corrupted by multiplicative noise with $L=2$ (the first column) and $L=4$ (the second column). From top to bottom: Noisy images (9.28 dB, 12.28 dB), DZ (21.80 dB, 23.01 dB), TwL-4V (22.57 dB, 23.68 dB), Ⅰ-DIV (22.39 dB, 23.48 dB), HMNZ (22.63 dB, 24.03 dB), and Ours (22.76 dB, 23.98 dB)

Download: Full-size image PowerPoint slide

Algorithm 1 Fixed-point algorithm based on the proximity operators for model (3).
Input: noisy image $f>0$ in $\mathbb{R}^{n}$; parameters $\lambda>0$, $\mu$, $\beta>1$; $\alpha>0$,
Initialization: $x^{(0)}$ and $y^{(0)}=0$; positive numbers $\sigma$ and $\rho$ such $\mu<\sigma$ and $\frac{\rho}{\mu}>8\sin^2\frac{(\sqrt{n}-1)\pi}{2\sqrt{n}}$.
repeat
(a) $x^{(k+1)}\leftarrow\mathrm{prox}_{\frac{1}{\rho}\Phi}(x^{(k)}-\frac{\mu}{\rho}H^\top(Hx^{(k)}-y^{(k)}))$,
(b) $y^{(k+1)}\leftarrow\mathrm{prox}_{\frac{\lambda}{\sigma}\psi}(y^{(k)}+\frac{\mu}{\sigma}(Hx^{(k+1)}-y^{(k)}))$,
until converges or satisfies a stopping criteria.
Write the output of $x^{(k+1)}$ from the above iteration as $\overline{x}$.
The restored image is $u^\star=e^{\overline{x}}$.

| Show Table

DownLoad: CSV

Table 1. Parameter values in our algorithm (Algorithm 1) at various noise levels

	$\lambda$	$\alpha$	$\beta$	$\mu$	$\rho$	$\sigma$
$L=10$	0.306	0.0015	6.06	30	250	150
$L=6$	0.406	0.00085	15.02	30	250	150
$L=4$	0.506	0.000108	250	30	255.5	90.26
$L=2$	0.8	0.00001	1655.05	30	290	156.26

| Show Table

DownLoad: CSV

Table 2. Parameter values for all testing algorithms

$L$	Method	$\lambda$	$\alpha$	$\beta$	$\mu$	$\rho$	$\sigma$
10	Ours	0.569	1.0002	1.00053	20.1	425.5	20.140
	TwL-4V	3.6/$L$	1.0	$-$	$-$	0.3	$-$
	Ⅰ-DIV	0.31	378.0	$-$	$-$	$-$	$-$
	DZ	0.07	19	3.0	3.0	$-$	$-$
	HMNZ	0.1	$-$	10	17.5	$-$	$-$
6	Ours	0.555	0.3885	1.005	19.6	285.5	19.667
	TwL-4V	2.9/$L$	1.0	$-$	$-$	0.29	$-$
	Ⅰ-DIV	0.45	918.0	$-$	$-$	$-$	$-$
	DZ	0.06	$3.8$	$3.0$	$3.0$	$-$	$-$
	HMNZ	0.1	$-$	10	9	$-$	$-$
4	Ours	0.659	0.18515	1.0105	29.658	255.5	29.918
	TwL-4V	2.4/$L$	1.0	$-$	$-$	0.3
	Ⅰ-DIV	0.55	658.0	$-$	$-$	$-$	$-$
	DZ	0.05	$1.59$	$3.0$	$3.0$	$-$	$-$
	HMNZ	0.1	$-$	10	6	$-$	$-$
2	Ours	0.8	0.000001	115.0	21.0	168	26.26
	TwL-4V	1.8/$L$	1.0	$-$	$-$	0.3	$-$
	Ⅰ-DIV	0.84	1059.0	$-$	$-$	$-$	$-$
	DZ	0.065	$0.45$	$3.0$	$3.0$	$-$	$-$
	HMNZ	0.1	$-$	10	1.5	$-$	$-$

| Show Table

DownLoad: CSV

Table 3. PSNR (dB) and CPU time (s) for Ⅰ-DIV[29], DZ[11], TwL-4V[17], HMNZ[14], our algorithm (Algorithm 1 with $x^{(0)}=\log(\text{SAR-BM3D}(f))$), and SAR-BM3D[25] for test images of Fig. 1 corrupted by multiplicative noise with $L=10, 6, 4, 2$, respectively

Image	$L$		Noisy	Ⅰ-DIV	DZ	TwL-4V	HMNZ	Ours	SAR-BM3D
Camer.	10	PSNR	15.61	28.69	28.30	28.83	29.96	30.15	28.89
	10	Time	$-$	6.28	86.30	6.23	66.62	117.35+6.03	117.35
	6	PSNR	13.39	27.35	26.93	27.59	28.46	28.53	26.35
	6	Time	$-$	11.40	110.36	7.33	62.86	116.59+10.11	116.59
	4	PSNR	11.64	26.52	25.85	26.72	27.33	27.38	23.67
	4	Time	$-$	13.57	149.54	7.94	63.69	118.31+13.31	118.31
	2	PSNR	8.63	24.98	24.27	25.28	25.30	25.47	16.45
	2	Time	$-$	15.57	190.75	9.72	62.58	124.30+15.20	124.30
Lena	10	PSNR	15.64	28.47	27.51	28.60	29.41	29.65	28.48
	10	Time	$-$	6.57	85.90	6.85	65.15	118.64+6.33	118.64
	6	PSNR	13.42	27.34	26.24	27.48	28.07	28.14	25.92
	6	Time	$-$	12.06	106.87	7.29	63.69	116.62+10.42	116.62
	4	PSNR	11.68	26.64	25.45	26.72	27.01	27.37	23.63
	4	Time	$-$	13.78	144.34	8.48	63.03	118.32+13.14	118.32
	2	PSNR	8.71	25.07	24.01	25.17	25.21	25.50	16.46
	2	Time	$-$	17.16	181.92	10.73	62.88	124.56+15.71	124.56
Pepp.	10	PSNR	15.93	28.83	27.20	28.86	29.13	29.53	28.09
	10	Time	$-$	6.88	84.15	7.47	65.33	116.86+5.62	116.86
	6	PSNR	13.70	27.95	26.15	27.92	28.12	28.51	25.78
	6	Time	$-$	12.46	108.44	8.16	60.54	118.31+10.24	118.31
	4	PSNR	11.98	27.10	25.10	27.05	27.14	27.76	23.63
	4	Time	$-$	13.30	145.13	8.54	63.75	119.16+12.90	119.16
	2	PSNR	8.93	25.57	23.72	25.54	25.52	25.97	16.45
	2	Time	$-$	17.44	188.50	10.59	61.59	115.71+15.03	115.71
Rem.1	10	PSNR	16.27	25.22	25.15	25.33	26.16	26.21	25.43
	10	Time	$-$	12.94	190.94	12.78	108.13	210.80+11.14	210.80
	6	PSNR	14.00	24.17	24.01	24.41	25.02	25.07	23.64
	6	Time	$-$	23.09	222.15	13.63	102.92	211.02+23.01	211.02
	4	PSNR	12.28	23.48	23.01	23.68	24.03	23.98	22.20
	4	Time	$-$	27.56	286.88	14.95	105.97	211.00+26.75	211.00
	2	PSNR	9.28	22.39	21.80	22.57	22.63	22.76	16.45
	2	Time	$-$	32.09	347.21	17.75	99.72	211.89+28.51	211.89
Rem.2	10	PSNR	16.23	25.56	25.59	25.63	26.76	26.69	25.92
	10	Time	$-$	9.16	144.40	9.58	84.96	163.00+8.90	163.00
	6	PSNR	14.02	24.55	24.30	24.45	25.25	25.43	24.07
	6	Time	$-$	17.46	164.92	9.94	85.24	162.70+17.03	162.70
	4	PSNR	12.27	23.45	23.38	23.63	23.93	24.26	22.26
	4	Time	$-$	22.75	219.99	11.79	78.15	163.07+20.35	163.07
	2	PSNR	9.22	22.01	21.86	22.18	22.12	22.72	16.08
	2	Time	$-$	25.86	271.57	14.01	75.17	162.84+24.02	162.84

| Show Table

DownLoad: CSV

Related Papers

Cited by

References

[1]	G. Aubert and J. Aujol, A variational approach to removing multiplicative noise, SIAM J. Appl. Math., 68 (2008), 925-946. doi: 10.1137/060671814.
[2]	R. Bamler, Principles of synthetic aperture radar, Surv. Geophys., 21 (2000), 147-157.
[3]	H. L. Bauschke and P. L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, AMS Books in Mathematics, Springer New York, 2011. doi: 10.1007/978-1-4419-9467-7.
[4]	J. M. Bioucas-Dias and M. A. T. Figueiredo, Multiplicative noise removal using variable splitting and constrained optimization, IEEE Trans. Image Process., 19 (2010), 1720-1730. doi: 10.1109/TIP.2010.2045029.
[5]	M. F. C. Chesneau and J. Starck, Stein block thresholding for image denoising, Appl. Computat. Harmon. Anal., 28 (2010), 67-88. doi: 10.1016/j.acha.2009.07.003.
[6]	K. Dabov, A. Foi, V. Katkovnik and K. Egiazarian, Image denoising by sparse 3d transform-domain collaborative filtering, IEEE Trans. Image Process., 16 (2007), 2080-2095. doi: 10.1109/TIP.2007.901238.
[7]	D. Dai, L. Shen, Y. Xu and N. Zhang, Noisy 1-bit compressive sensing: Models and algorithms, Appl. Computat. Harmon. Anal., 40 (2016), 1-32. doi: 10.1016/j.acha.2014.12.001.
[8]	I. Daubechies, Ten Lectures on Wavelets, vol. 61 of CBMS Conf. Series Appl. Math., SIAM, Philadelphia, 1992. doi: 10.1137/1.9781611970104.
[9]	I. Daubechies, B. Han, A. Ron and Z. Shen, Framelets: MRA-based constructions of wavelet frames, Appl. Comput. Harmon. Anal., 14 (2003), 1-46. doi: 10.1016/S1063-5203(02)00511-0.
[10]	L. Denis, F. Tupin, J. Darbon and M. Sigelle, SAR image regularization with fast approximation discrete minimization, IEEE Trans. Image Process., 18 (2009), 1588-1600. doi: 10.1109/TIP.2009.2019302.
[11]	Y. Dong and T. Zeng, A convex variational model for restoring blurred images with multiplicative noise, SIAM J. Imag. Sci., 6 (2013), 1598-1625. doi: 10.1137/120870621.
[12]	S. Durand, J. Fadili and M. Nikolova, Multiplicative noise removal using l1 fidelity on frame coefficients, J. Math. Imag. Vis., 36 (2010), 201-226.
[13]	J. W. Goodman, Some fundamental properties of speckle, J. Opt. Soc. of Amer., 66 (1976), 1145-1150. doi: 10.1364/JOSA.66.001145.
[14]	Y. Hang, L. Moisan, M. K. Ng and T. Zeng, Multiplicative noise removal via a learned dictionary, IEEE Trans. Image Process., 21 (2012), 4534-4543. doi: 10.1109/TIP.2012.2205007.
[15]	Y.-H. Huang, H. Yan and T. Zeng, Multiplicative noise removal based on unbiased box-cox transformation, Communications in Computational Physics, 22 (2017), 803-828. doi: 10.4208/cicp.OA-2016-0074.
[16]	Y.-M. Huang, M. K. Ng and Y.-W. Wen, A new total variation method for multiplicative noise removal, SIAM J. Imag. Sci., 2 (2009), 20-40. doi: 10.1137/080712593.
[17]	M. Kang, S. Yun and H. Woo, Two-level convex relaxed variational model for multiplicative denoising, SIAM J. Imag. Sci., 6 (2013), 875-903. doi: 10.1137/11086077X.
[18]	D. Lazard, Quantifier elimination: Optimal solution for two classical examples, J. Symbol. Comput., 5 (1988), 261-266. doi: 10.1016/S0747-7171(88)80015-4.
[19]	F. Li, M. K. Ng and C. Shen, Multiplicative noise removal with spatially varying regularization parameters, SIAM J. Imag. Sci., 3 (2010), 1-20. doi: 10.1137/090748421.
[20]	J. Lu, L. Shen, C. Xu and Y. Xu, Multiplicative noise removal in imaging: An exp-model and its fixed-point proximity algorithm, Appl. Comput. Harmon. Anal., 41 (2016), 518-539. doi: 10.1016/j.acha.2015.10.003.
[21]	C. A. Micchelli, L. Shen and Y. Xu, Proximity algorithms for image models: Denoising Inverse Probl., 27 (2011), 045009(30pp). doi: 10.1088/0266-5611/27/4/045009.
[22]	J.-J. Moreau, Fonctions convexes duales et points proximaux dans un espace hilbertien, C.R. Acad. Sci. Paris Sér. A Math., 255 (1962), 2897-2899.
[23]	Y. Nesterov, Introductory Lectures on Convex Optimization, Kluwer, Boston, 2004. doi: 10.1007/978-1-4419-8853-9.
[24]	C. Oliver and S. Quegan, Understanding Synthetic Aperture Radar Imaging, SciTech Publishing, Raleigh, NC, 2004.
[25]	S. Parrilli, M. Poderico, C. V. Angelino and L. Verdoliva, A nonlocal SAR image denoising algorithm based on LLMMSE wavelet shrinkage, IEEE Trans. Geosci. Remote Sens., 50 (2012), 606-616. doi: 10.1109/TGRS.2011.2161586.
[26]	L. Rudin, S. Osher and E. Fatemi, Nonlinear total variation based noise removal algorithms, Phys. D, 60 (1992), 259-268. doi: 10.1016/0167-2789(92)90242-F.
[27]	J. M. Schmitt, S. H. Xiang and K. M. Yung, Speckle in optical coherence tomography, J. Biomed. Opt., 4 (1999), 95-105. doi: 10.1117/1.429925.
[28]	J. Shi and S. Osher, A nonlinear inverse scale space method for a convex multiplicative noise removal, SIAM J. Imag. Sci., 1 (2008), 294-321. doi: 10.1137/070689954.
[29]	G. Steidl and T. Teuber, Removing multiplicative noise by Douglad-Rachford splitting methods, J. Math. Imag. Vis., 36 (2010), 168-184. doi: 10.1007/s10851-009-0179-5.
[30]	T. Teuber and A. Lang, Nonlocal filters for removing multiplicative noise, Proc. of SSVM, LNCS, 6667 (2012), 50-61. doi: 10.1007/978-3-642-24785-9_5.
[31]	R. F. Wagner, S. W. Smith, J. M. Sandrik and H. Lopez, Statistics of speckle in ultrasound B-scans, IEEE Trans. Sonics and Ultrason., 30 (1983), 156-163. doi: 10.1109/T-SU.1983.31404.
[32]	S. Yun and H. Woo, A new multiplicative denoising variational model based on m-th root transformation, IEEE Trans. Image Process., 21 (2012), 2523-2533. doi: 10.1109/TIP.2012.2185942.