International Journal of Advances in Science and Technology, Vol. 3, No.

3, 2011

Denoising Algorithms for Vector-Valued Images Using GCM Models

*1, D.Kartheek and #2 J.Brahmaih naik, #3 M.Venugopalarao, *4 V.Ajesh Kumar *1 Post graduate student, Narasaraopet Engg.COLLEGE, Narasaraopet, A.P., India karthik.dasari75@gmail.com #2 Assistant professor, Dept .of ECE, Narasaraopet engineering college, Narasaropet, A.P., India brahmaiah05@yahoo.co.in #3 Professor & HOD, Dept. of ECE, Narasaraopet Engg.COLLEGE, Narasaraopet, A.P., India mvgr03@gmail.com *4 Post graduate student, Sir C.R.REDDY Engg.college, Eluru, A.P., India ajeshkumar.palakol@gmail.com
Abstract

Variational techniques for gray-scale image, color images de-noising have been deeply investigated for many years; however, little research has been done for the vector-valued denoising case and the very few existent works are all based on total-variation regularization. In this paper we would like to introduce High-order models for vector-valued images. Their properties are analyzed and a fast multigrid algorithm for the numerical solution is provided. In this paper we introduce Higher Order models corresponding to their algorithms. By using two different methods we can justify the denoised images whether noise was removed or not. By seeing PSNR value.
Index TermsFourth-order partial differential equations (PDEs), image denoising, multilevel methods.

I.

INTRODUCTION

DENOISING of gray-scale images has been extensively studied and investigated within the last decades. With the appearance of the total variation (TV) model of Rudin, Osher,and Fatemi [23], it became evident that variational approaches to the image denoising problem can yield often excellent results. However, for images with smooth features and non piecewise constant intensities, the TV model may produce images appearing blocky with the so called staircase effect. There exist several ways of improving the model by a different regularize (see [20],[27] and [10] and refence there in) and one effective approach is to use high-order models of partial differential equations (PDEs). This paper will address such types of high-order models for vector valued images (including color images with three channels) and the associated fast algorithms. Surprisingly and up to our knowledge, there is however no published work of such a kind of vectorvalued high-order model. Of course, we expect such a model to deliver better results than its counterpart (the high-order CbC approach). In this paper, we take the high-order and curvature-based denoising model of Zhu-Chan [35] for gray-scale images as our starting point and introduce two different ways to generalize it to vectorvalued images obtaining respectively what we call global and local coupling among the channels. We will analyze the properties of these two new models and will show some examples suggesting that our global high-order model is not only better than its correspondent CbC high-order competitor but also better than TV based color models. Finally, we will show how to implement a fast multigrid algorithm for this recommended model.

Special Issue I

Page 66 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

II. PROBLEM FORMULATION AND CURRENT MODELS Define a vector-valued image as a function i.e., with . A noisy image is obtained by adding up Gaussian noise to i.e., . The variational approach to remove is then (1) Where selecting the space of functions were will belong. Red Green Blue = and is a regularization term

Fig. 1. Model problem, an image with a weak piecewise constant red channel, a piecewise smooth green channel, and a strong piecewise constant blue channel.

We will use this image to test the models we review here. We now proceed to present in historical order three different ways to generalize the total variation model of Rudin, Osher, and Fatemi [23] from gray-scale to vector-valued images. This will prove to be helpful at the moment of introducing our highorder color models. III. NEW AND HIGH-ORDER VECTOR-VALUED MODELS We are ready to state our proposed models. As in the vectorial TV case we need to have a starting point and from there start moving on. In other words, we have to select a working high-order model for gray-scale images and upgrade it to vector valued images. The idea is also to have coupling among channels as in the vectorial TV models since this has proven to improve the quality of reconstruction. To this end we decided to use the curvature-based model [35] as our starting point. Our motivations are three fold: 1) this model has nice properties like no staircase effect ad contrast and corners preservation [35], 2) a fast multigrid solver can be implemented for this model, and 3) curvature is an intrinsic geometric feature so generalization is easier. First we review very briefly what it would be the CU1-CbC approach of this curvature-based model for color image denoising. A. Channel by Channel Curvature-Based Model(CU1-CbC) Denote as usual the true vector-valued image, the noisy image =(), and the curvature vector with the curvature of the th channel of . Then solve which may be done by solving the system of EulerLagrange equations in for l=1,2,.m : with boundary conditions for eachchannel. The objective of is to efficiently measuring therandom high frequency oscillations representing noise in . Many different high-order options including or have been studied; some of those can be found in [34], [18], [19], and [7] and references therein. In particular, we use here and throughout the rest of this work due to its nice properties already summarized above and detailed in [35].With such a selection .

Special Issue I

Page 67 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

Here as in the TV1-CbC approach for the vectorial TV model,is the same for all channels so we expect to have difficulties when denoising a color image having weak channels. This is, by selecting the best for one channel we may over-smooth the weak channel. We illustrate this effect in the Fig.1. B. Local Curvature-Based Color Model(CU2-LCM) Here we introduce our first high-order denoising model for vector-valued images. We construct this model based on ideas from the Bresson and Chan model. This model is to solve

(3) which leads to solve the EulerLagrange equations (4) in for , with boundary conditions and defined as

(5) As can be appreciated, the amount of diffusion in this model, is mainly affected by , a vector that locally varies across the image. Due to this, we name this model the local curvature based model (CU2-LCM). Red Green Blue

Fig.2. Result from the CU2-LCM model. The small step in the weak red channel is better recovered but a kink appeared in the blue channel.

Unfortunately, our experiments revealed that this model has a tendency to develop kinks (see the blue channel in Fig. 2) in the image not only decimating the quality of restoration but making more difficult to construct stable numerical solvers as well. C. Global Curvature-Based Color Model(CU3-GCM) Our second new high-order model is inspired by the Blomgren and Chan TV-based model. Here we propose to solve

(6) which leads to the EulerLagrange equations in (7) with boundary conditions for l=1,m where Q(u)and Q1(ul) are defined by

Special Issue I

Page 68 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

(8) Note that in the total variation TV3-BLC model, the amount of regularization or diffusion is determined by the quantity while here with the curvaturemodel CU3GCM we have which acts as a regularization weight for channel . For this reason, we call this model the global curvature-based model. Red Green Blue

Fig. 3. Result from the CU3-GCM model. All channels, including the weak one, are very well reconstructed.

IV. NUMERICAL SOLUTION OF THE HIGH-ORDER MODELS A. Stable Fixed Point Method Here we present a general method to implement a fixed point algorithm for any of the high-order models presented so far. Later on this section, we will show how to modify slightly this method to obtain an optimal performance for the CU3-GCM model. Then we will introduce a nonlinear multigrid method for it. The EulerLagrange equations of all of the high-order models can be written in the general form (9) where for and, respectively, for the CU1-CbC model, the CU2-LCM model, and the CU3-GCM model

To obtain a fast solution of the above PDE a fixed point ethod similar to the ones described in [33], [32], [8], and [26] for the TV model would be desirable. Straight forward implementation of such a FP method

Special Issue I

Page 69 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

does not work for PDEs like (9) so we use a method where a stabilizing term N() is included and the following fast fixed point scheme is used:

(10)

For the CU1-CbC and CU3-GCM models is selected. For the CU2-LCM model however provides a better performance of the algorithm. _________________ Algorithm 1 CFPGS Require: On a grid with mesh size , choose an initial guess = () for (10) 1: for to do 2: for to do 3: Apply Gauss-Seidel iterations to the linear System 4: end for 5: end for From the evidence shown in Section III, there is clear indication that the CU3-GCM method is the best among the high-order models. Due to this, we now concentrate on optimizing the Algorithm 1 for this model. B. Nonlinear Multigrid Algorithm We can go one step further and use the above fixed-point method as the foundation for a nonlinear FAS MG algorithm [31], [9]. This algorithm has been successfully developed and tested in a number of imaging problems, for instance: [15] on image registration, [1], [22] on image segmentation, [26], [27], [6], [13], [11] on image denoising-deblurring, and [5] on image in painting. Multigrid schemes considerably speed up numerical processes achieving fast results by constructing a hierarchy of discretizations where at each level the error equation is partially solved and the new approximation transported to next coarser level. This process is recursively applied until reaching the coarsest level where and exact but computationally cheap solution is obtained. Then the process move backwards on the hierarchical structure transporting the more accurate error and updating the approximate solution at each level until reaching the finest level again. Usually standard coarsening is used to construct the hierarchical structure halving the number of variables on each dimension at each level. To apply this scheme to our problem we define the nonlinear discrete equations

where , and define such that as the vector of nonlinear operators on the grid of mesh size h . Define the residual equations as and the

Special Issue I

Page 70 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

correspondent vector .Then the nonlinear MG scheme for the vector-valued problem is stated in Algorithms 2 and 3. __________________________________ Algorithm 2 Nonlinear Multigrid Method __________________________________ Require: Select an initial guess on the finest grid 1: 2: while do 3: 4: 5: 6: end while ____________________________ Algorithm 3 FAS Cycle ____________________________

1: if =coarsest grid then accurately (i.e., iterations by CFPGS) and return. 2: solve 3: else 4: continue with step 6. 5: end if 6: Pre-smoothing: Do steps of 7: Restrict to the coarse grid, 8: Set the initial solution for the next level, 9: Compute the new right hand side 10: Implement 11: Add the residual correction, 12: Post-smoothing: Do steps of

As a smoother we use our fixed-point CFPGS algorithm; however, to achieve the best possible performance of the MG algorithm, we found that applying over-relaxation into the GS sweeps (SOR method [24]) considerably speeds convergence, i.e., we update the unknown using with = 1.5. We also apply a small number of extra local relaxation steps around the difficult points (edges of the image) as suggested in [3] which is where the most inhomogeneous parts are located. This guarantees the residual to be smooth enough before being transported to the next coarser grid and is computationally very cheap since these regions represent normally less than 1/7 of the whole image domain so the overall cost of the MG algorithm is incremented in a very small percentage. V. FURTHER NUMERICAL EXPERIMENTS We now proceed to show some results obtained using the CU3-GCM model and the multigrid algorithm outlined above, using three realistic color images.

Special Issue I

Page 71 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

Of the three high-order models (CU1-CbC, CU2-LCM, and CU3-GCM) we have presented here, our experiments suggest that CU3-GCM is better than the other two. Due to these facts, we only compare the quality of restoration yielded by the CU3-GCM models.

(b)

(c)

Fig.5. Residuals versus the cycles of a nonlinear multigrid algorithm when solving the problems.

(a)

(b)

Fig.6. Example 2: (a) Denoising example using the CU3-GCM model. (b) Denoising example using the TV3-BLC model. (c) A magnified region taken from (a). Correspondent magnified region taken from (b)

(b)

Fig.7. Example 3: (a) Denoising example using the CU3-GCM model. (b) Denoising example using the TV3-BLC model.

the quality of restoration between TV3-BLC and CU3-GCM. Although both restorations of Fig. 13(a) and (b) may look very similar, when details are magnified as in Fig. 13(c) and (d), the CU3-GCM model shows a superior quality of restoration. Similar behavior is observed in Fig. 7. B. Multigrid Performance Now we proceed to illustrate the fast performance of the nonlinear multigrid algorithm for the CU3GCM model. In Fig.5

Special Issue I

Page 72 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011

TABLE I
COMPARISON OF CPU-TIMES AND PSNR VALUES FROM THE MG AND CFPGS ALGORITHMS WHEN SOLVING THE PROBLEMS OF FIGS. 1214. THE MG ALGORITHM IS ROUGHLY FROM 5 TO 6 TIMES FASTER AND BOTH DELIVER THE SAME QUALITY OF RECONSTRUCTION

we present the history iteration for solving each one of the problems from Figs. 5-7 all with SNR20 . For these problems we used the following parameters=10-2, = 2.Clearly, the MG iteration is very good reaching very quickly very small residuals. A good stopping criteria for the MG algorithm is to stop when the relative residual is less than 10-4. In Table I, we present the CPU-time consumed for both, MG and CFPGS algorithms, when solving the same problems of above now with noisier images i.e., SNR and using the stopping criteria just described above. Although both methods are pretty fast, MG is roughly from 5 to 6 times faster than CFPGS.We also present the obtained PSNR values to illustrate the well-balanced restoration of all of the three channels of the color images. All simulations were carried out using Matlab 2008a on a 2.8-GHz IntelXeon-based computer. VI. CONCLUSION The new models are designed to have coupling among the channels of the color image which many researchers have identified as highly desirable. Of the presented three curvature models, the global CU3GCM model delivers the best results. Further a fast numerical multigrid algorithm is constructed for this model. A comparison between the new CU3-GCM model and the previously best TV3-BLC model was presented in denoising color images. The CU3-GCM model is found to cope better with unbalanced channels in the common situation where different levels of illumination are present when a color picture (with noise) is produced.
REFERENCES [1] N. Badshah and K. Chen, Multigrid method for the chan-vese model in variational segmentation, Commun. Comput. Phys., vol. 4, no. 2,pp. 294316, 2008. [2] P. Blomgren and T. F. Chan, Color TV: total variation methods for restoration of vector-valued images, IEEE Trans. Image Process., vol.7, no. 3, pp. 304309, Mar. 1998. [3] A. Brandt, Multigrid techniques: 1984 guide with applications to fluid dynamics, Gesellschaft fr Mathematik und Datenverarbeitung, 1984,GMD-Studie Nr. 85.

Special Issue I

Page 73 of 74

ISSN 2229 5216

International Journal of Advances in Science and Technology, Vol. 3, No.3, 2011 [4] X. Bresson and T. F. Chan, Fast dual minimization of the vectorial total variation norm and applications to color image processing, InverseProblems and Imaging, vol. 2, no. 4, pp. 455484, 2008. [5] C. Brito-Loeza and K. Chen, Multigrid method for a modified curvature driven diffusion model for image inpainting, J. Comput. Math.,vol. 26, pp. 856875, 2008. [6] T. F. Chan, K. Chen, and J. L. Carter, Iterative methods for solving the dual formulation arising from image restoration, Electron. Trans.Numer. Anal., vol. 26, pp. 299311, 2007. [7] T. F. Chan, A. Marquina, and P. Mulet, High-order total variationbased image restoration, SIAM J. Sci. Comput., vol. 22, no. 2, pp.503516, 2000. [10] Y. Chen, S. Levine, and M. Rao, Variable exponent, linear growth functionals in image restoration, SIAM J. Appl. Math., vol. 66, no. 4,pp. 13831406, 2006. [11] M. Donatelli, A multigrid for image deblurring with Tikhonov regularization, Numer. Linear Algebra With Applicat., vol. 12, pp.715729, 2005. [12] M. Do Carmo, Differential Geometry of Curves and Surfaces.. New York: Prentice-Hall, 1976. [13] C. Frohn-Schauf, S. Henn, and K. Witsch, Nonlinear multigrid methods for total variation image denoising, Comput. Visualiz. inSci., vol. 7, pp. 199206, 2004. [14] S. Henn and K. Witsch, A multigrid approach for minimizing a nonlinear functional for digital image matching, Computing, vol. 64, pp.339348, 2000. [15] L. Homke, A multigrid method for anisotropic PDEs in elastic imageregistration, Numer. Linear Algebra With Applicat., vol. 13, pp.215229, 2006. [35] W. Zhu and T. F. Chan, Image denoising using mean curvature, [Online].Available: http://www.math.nyu.edu/~wzhu/, Preprint
1 Mr. D.Kartheek is a postgraduate student, department of ECE in Narasaraopet Engineering College, Narasaraopet, India.Now he is pursuing on his post graduation studies i.e. M.Tech (DIGITAL SYSTEMS AND COMPUTER ELECTRONICS).he is interested to do research in Image Processing.

2 Mr.J.Brahmaiah naik working as assistant professor in Narasaraopet Engineering College, Narasaraopet, India. He obtained M.Tech degree from JNTUnivesity, KAKINADA. He is interested do research in Image processing, communication systems.

3. Sri M.venu gopala rao, working as professor & HOD in department of ECE, Narasaraopet Engineering COLLEGE, Narasaropet, India. He is interested to do research in Image processing. 4. V.Ajesh Kumar is a post graduate student in Sir C.R.Reddy College of engineering, Eluru, A.P., india.His area of interest is Turbo sciences.

Special Issue I

Page 74 of 74

ISSN 2229 5216