非线性反应—扩散方程在图像处理中的若干应用
摘要
本文主要研究非线性反应扩散方程在图像恢复、分解和分割中的若干应用.
在第一章里,我们提出两个用于图像去噪和分解的反应扩散方程组模型.在第一部分中,受到Osher等人提出的基于H-1模的图像恢复和分解泛函模型的启发,我们提出一个耦合的反应扩散方程组模型.新模型包含两个主部不同的方程：其中一个主部为全变差扩散,用来去除噪声；另外一个主部为热扩散,用来修正去噪方程的源项,达到保护图像纹理的效果.两个方程相互作用,最终达到去噪的同时又能很好的保持边界和纹理.在第二部分中,在上述方程组模型的框架下,我们引入p(x)-Laplace流来代替原来的全变差流,从而新的去噪方程结合Gaussian热方程的各向同性扩散和全变差流的各向异性扩散的特点,达到自适应的效果.在理论研究方面,利用正则化方法以及Galerkin方法我们分别得到两个模型解的适定性,这为数值计算提供了必要的准备.与原来的模型不同,新模型基于反应扩散方程组框架.另外,新模型在去除阶梯效应、保持边界和去噪方面都有很好的效果,尤其在保持纹理方面效果显著.
在第二章里,我们提出用于图像去噪的混合型扩散方程模型.通过分析图像边界与图像位置和结构的依赖关系,构造出一个抗噪能力较强的边界映射函数,对应的扩散方程不仅具有各向异性扩散的特点,而且在图像内部齐次区域进行Gaussian热扩散,在近边界区域进行平均曲率扩散,最终达到自适应去噪的效果.在理论研究方面,我们首先利用不动点方法证明新模型的适定性,随后研究t→∞时解的渐近状态,结论表明新模型恢复结果的长时间极限状态为初始图像的局部均值.最后,我们提出一种计算格式,分析其收敛性,并对一系列图像进行数值实验.与TV模型相比,新模型在内部齐次区域进行Gaussian热扩散,有效避免阶梯效应；与PM模型相比,新模型是适定的,从而解更加稳定,而且由于新模型的扩散模式和扩散速度依赖于图像的位置和结构信息,模型自适应能力更强,在内部齐次区域能更好的去噪,在近边界处能更精细的保持边界.
在第三章中,我们基于活动轮廓方法建立两个图像分割模型.第一个模型主要用于加快基于活动轮廓的水平集分割算法.通过分析CV无边界活动轮廓水平集分割算法对光滑度不同的图像的分割结果,我们提出一种新的图像分割快速算法.新算法去掉原模型的正则项,在灰度水平集上离散泛函能量进行计算,计算复杂度为O(N),计算时间很短,而且分割结果和原始模型相近.第二个模型主要是改进彩色图像的分割效果.基于偏微分方程彩色图像去噪模型,我们提出一个凸的能量泛函向量值图像分割模型,并根据对偶原理,构造出求解新模型的算法.数值实验显示新算法的分割结果更加符合实际要求,而且新算法也相对快速稳定.
The image is between the humanity exchanges one of most direct-viewing ways. Especially in current information technology rapid development time, image infor-mation as an important communication medium are widely used in multimedia, digital health, artificial intelligence, aerospace and geological remote sensing, and other fields. Generally speaking, the digital image processing technology take the digital image as an object, including image acquisition, image analysis and image understanding three main areas. Image analysis is between image acquisition and image understanding between one of the links, it directly affects computer vision, robot vision, and machine vision. In image processing, image restoration and image segmentation is important problem and has been extensively studied, they all belong to the image analysis of research areas.
Image processing research has a long history, goes back to the 1920s, most of the image processing methods are based on a dimensional signal processing technologies, such as filtering technology, statistical theory, etc. With remote sensing, medical, and other fields growing demands, as well as physical technology breakthrough, peo-ple are increasingly concerned about the nature of image processing, and attempt to use strict mathematical theory to the current image processing method for colla-tion and improvement. Currently, the image processing technology has three main tools:Stochastic Models, Wavelet Analysis Theory and Partial Differential Equa-tions Methods. The first two have a long history:Stochastic Models are based on Bayes and markov random field and are to work directly on the images; Wavelet Analysis Theory is based on Fourier analysis, transform the image information to the frequency domain information, and then are used for image processing. Since the 1990s, image processing methods based on PDEs (partial differential equations) developed rather quickly, becoming a very attractive study areas. At present, in im-age restoration, image segmentation, image reconstruction, image recognition, image analysis and so on, image processing methods based on PDEs has been widely ap-plied. In PDEs models for image processing, image information is quantified as a continuous two-dimensional function, and can be seen as a field or physical state. PDEs methods work directly on the image, is closer to reality the objective world and easy to analysis models. In the mathematical theory, there are lots of the the-ories on PDEs, and then the models can be deeply analyzed. From this stage of development, we can not know which is the best in these three tools, And these three tools is kept in touch with each other and cross-promotion.
In this article, the first problem we are concerning is PDEs methods for image restoration and decomposition and the corresponding theoretical analysis. In image processing, image restoration has always been one of the most important fundamen-tal research subjects. In the image acquisition, copy, scan, transmit, display, and so on,the images will be inevitably degraded, such as blur and noise. However, in many application areas, people need a clear, high-quality images, so image restora-tion (such as denoising, deblurring, etc) is of great significance. In theory, noise is unpredictable random error, and subject to some probability distribution. Assume that u:Ω(?) R2→R is the intensity of the ideal image, n(x) is the noise,f is the
intensity of the observed image. The idea image u and the observed image f are
related by
f= u+n.
Generally speaking, the noise n(x) satisfies:
where, the first constrain indicates that the noise has zero mean, and the second one uses a priori information that the standard deviation of the noise n(x) isη. For ease of analysis, we assume that the image signal and noise are.independent of each other. The problem is to recover the restoration image, u, from the observed, noisy image, u0, where the two are related by u0= u+n(χ). The image restoration and decomposition models based on PDEs have the two important framework:the variational methods based on the optimal problem for energy functionals and the diffusion equation methods based on fluid diffusion theory. Image decomposition is the extension of image noise, Image denoising model basically can also be used as image decomposition model. Next, we present some image denoising models based on the variational method, image decomposition models based on the variational method, and image restoration models based on diffusion equations.
First, we introduce image denoising models based on the variational method and some classical denoising models. The famous classical and general method for removing Gaussian noise firs consists in solving the following constrained minimiza-tion problem:
λ> 0 is a weight parameter. The first term in E(u) is a smoothing term for deniosing. The second term in E(u) measures the fidelity to the initial data. In
1977, in the model proposed by Tikhonov and Arsenin [1],(?)(s)= s2. And the solution space is
whereΩis a bounded open domain of-RN with the appropriate smooth boundary, n denotes the unit outward normal to the boundary (?)Ω, and then the constrained minimization problem has unique solution in W. However, the above-mentioned (?) punish the gradient too much and W1,2(Ω) is a smoother space, and therefore the recovery image is too smooth to preserve the image edge information. The real image maybe not be smooth, so someone try to find a better function (?), for example, (?)(s)=|s|p,1< p< 2, and then the corresponding solution space is W1,p(Ω). In 1992,Rudin, Osher and Fatemi [2] propose the famous total variation method, the ROF model. In the TV model,(?)(s)=|s|, and mathematically this is reasonable, since it is natural to study solutions of this problem in the space of functions of bounded variation, BV(Ω), allowing for discontinuities which are necessary for edge reconstruction. The TV model has been studied extensively and has proved to be an invaluable tool for preserving edges in image restoration problem. In 1997, Chambolle and Lions [3] detailed theoretical analysis for the TV model, and propose the corresponding numerical calculation method. In 2003, Rudin, Lions and Osher [4] summary theoretical results for the total variation model, as well as the related theory of partial differential equations. When (?)(s)= (1+s2)~(1/2), the corresponding model is based on the minimal surface functional, and the functional has similar properties with the TV model and well-posedness [13]. When (?)(s) is the non-convex function, such as(?)(s)= log(1+s2)[9], using the model, we can have better results for image restoration. However,the existence and uniqueness of the solutions for the model still are a open problem, the convergence of the numerical scheme for the model had not been proved. The TV model is well posed, but TV-based denoising favors the piecewise constant solutions. Sometimes this also causes a staircasing effect in which noisy smooth regions are processed into piecewise constant regions [3], [5], and [38]. Not only'blocky'solutions fail to satisfy the ubiquitous'eyeball norm' but they can also develop'false edges' which can mislead a human or computer into identifying erroneous features not present in the true image. Some authors consider another regularizing term to remove the noise [5], which is as follows
where (?) p(s)→2, (?) p(s)→1, and p is monotonically decreasing. This model should reap the benefits of both isotropic and TV-based diffusion, as well as a com-bination of the two. However, it is difficult to study mathematically since the lower semi-continuity of the functional is not readily evident. In [6], Chen, Levine, and Rao modify the model, and propose a functional with variable exponent,1≤p(x)≤2, which is a combination of Total Variation based regularization and Gaussian smooth-ing.
Secondly, we introduce some image decomposition model based on the varia-tional method. If we decompose the image into two components:the smoothed version of the original image f, denoted by u, which represents the cartoon in-formation of the original image, and the componentυ= f - u, which represents the texture or noise information, then the above model is actually a new energy functional for image restoration and image decomposition into cartoon plus texture, where the first term is a regularizing term to remove the noise, and the second term is a fidelity term. Though the ROF model performs very well for image denoising and edge protecting, it may also destroy small details, such as textures, see [7]. To overcome this drawback, Meyer [7] proposed a new minimization method by intro-ducing a weaker norm which is more appropriate to represent textured or oscillatory patterns. Subsequently, Osher et al. [8] presented the following model which com- bined the norm for oscillatory functions proposed by Meyer [7] (involving the H -1 norm) with the total variation minimization from the ROF model [2] whereλ> 0 is a weight parameter. Osher, Sole and Vese[8] is in turn a particular case of an earlier work by Osher and Vese [24]. Inspired by Meyer's G-norm, some authors have proposed various numerical models for image decomposition and mod-eling texture with Meyer's G-norm and F-norm, and also with Besov norms[25]-[27].
Finally, we present and analysis some denoising models based on the diffusion equations noise model, which is from the fluid diffusion physical phenomena. Early in the 1960s, the oldest and most investigated equation proposed by Gabor in the processing is probably the parabolic linear heat equation with the initial data f. Alvarez et al. [10] rigorously establish the connection between scale space analysis and PDEs. In [9], Perona and Malik developed an anisotropic diffusion scheme for image denosing. The basic idea of this nonlinear smoothing scheme was to smooth the image while preserving the edges in it. This was done by using the equation
where C(s) is a smooth non-increasing function such that C(s)→1, as s→0 and C(s)→0, as s→∞. In fact, the PM diffusion equation is a forward-backward diffusion and the well-posedness of the equation is still a open problem. It is worth mentioning that, the TV model can be seen as the Perona-Malik diffusion when c(|▽υ|)= (?). In [11], A simple adjustment with practical applications, is to include a short range mollifier in the nonlinear diffusion. Existence and uniqueness of solutions to this modified Perona-Malik equation has been proved by Catte et al. for initial dataυ0∈L2(Ω). Inspired by the PM model, many researchers proposed new models which is to work on the nonlinear anisotropic diffusion equations, without thinking of any energy functional. All models are nonlinear anisotropic diffusion filters with diffusivities being adapted to the local image structure. Weickert [12] propose the model which is a nonlinear anisotropic diffusion filter with diffusion tensors being adapted to the local image structure. The author modify the diffusion tensor matrix eigenvectors and eigenvalues to control the diffusion direction and diffusion strength. The models based on PDEs can easily extend to the color image case. In the color image case, the geometry properties of surfaces [13] is involved. In addition, the convection term in PDEs is introduced, and can speed up calculation speed [14]-[15]. In practice, the ill-posedness results in a mild instability in the PM problem. Regions of high gradients develop a "staircase" instability that involves dynamic coarsening of the steps as time evolves. To make the images more pleasing to the eye, it would be useful to reduce this effect. To this end, many researchers consider higher-order version of the Perona-Malik equations, examples of which.are given in [16]-[18].
In this article, the second problem we are concerning is image segmentation models based on PDEs and the level-sets method. Image segmentation is one of the most fundamental problems in the fields of image processing and computer vision. The aim is to find a partition of an image into its constituent parts. However, this definition is rather unsatisfactory and ambiguous. As we will see, the main difficulty is that one need to manipulate objects of different kinds:functions, domains in R2, and curves. In the last century in the late 1980s, partial differential equation method is first applied to the image segmentation. Kass, Witkin and Terzopoulos[19] first propose the snake and the geodesic active contours models. Mumford and Shah [20] propose the variational segmentation method for the region detection. Since the image edges is one-dimensional, in early segmentation algorithms it is difficult to indicate the edge information for images. At that time, using theΓ-convergence or spline curves and other techniques, the researchers obtain the image segmentation results. However, these methods are not only difficult to understand, and solve complex, and do not have the expected effect.1988年, Osher and Sethian [21] introduce the level-sets method which allows for cusps, corners, and automatic topological changes. The curve is represented implicitly via a Lipschitz function, by the zero level-set of the Lipschitz function, and the evolution of the curve is given by the zero-level curve from the evolution of the function. Since then, the PDEs method is widely used to image segmentation. The iconic work can be see the geodesic active contours models by Caselles, Kimmel and Sapiro[22] and the active contours without edges models by Chan and Vese [23]. At present, the CV model has many improvements and development and has attracted the attention of many scholars.
In this paper we studies image processing methods based on PDEs, especially in image restoration, decomposition and segmentation applications. Since in im-age processing the models based on PDEs (system) which is from the research and analysis of the image structure, is usually the new type diffusion equation (system), so the existence and uniqueness of solutions for the new PDEs(system) is impor-tant practical significance for the numerical calculation. In this paper the following problems are studies:the first chapter is the theoretical research and numerical implementation for the image restoration and decomposition model based on new partial differential systems; the second chapter is the theoretical research and nu-merical implementation for the image restoration model based on new anisotropic diffusion equations; the third chapter is the introduction and analysis of two new fast segmentation algorithms. every problem in this paper can be divided into the following sections:first, we introduce a new model, usually based on some diffu-sion equation (system); then, we have demonstrated the existence, uniqueness and boundedness of solutions for the new models; finally, we structure the numerical scheme which is applied to image processing with respect to other known models. Since the new equation (system) structure is different, we applied the regularization method, the Galerkin method, and the fixed-point method to obtain the existence of solutions for the new models, respectively. Then, experimental results illustrate the effectiveness of the new models with respect to other known models.
In the first chapter, the two image restoration and decomposition models based on reaction-diffusion PDEs (system) is proposed. In the first part of this chapter, inspired by Osher, Sole and Vese[8], we propose the following model based on the reaction-diffusion partial differential system: where f is the original image,Ωis a bounded domain of RN with appropriately smooth boundary, n the unit outward normal toΩ,1≤N< 4, T> 0,λ> 0, and as far as we know, there is few literature about PDE system devoted to image processing. Different from the models as before, in the new system, the two equations interact with each other:first we obtainωfrom (1) to adjust the fidelity betweenυand f in (2), then we use (2) which is from the classical TV model to obtain the smoother image u. In other words, the denoising equation (2) has a different fidelity termωwhich is also derived from an evolutionary equation, thus we can select an appropriate scale to achieve an appropriate fidelity betweenυand f and obtain a better result for noise removal and edge or texture preservation. Furthermore, we can proveυis the element of BV space theoretically, which allows for discontinuities along curves, therefore edges and contours are kept in the image u. In addition, in theory, the problem (2) can be seen as a TV flow equation with nonlinear source. In [28,29,30], Andreu et al. studied the TV flow ut= div(Du/|Du|) with L1-data using the Crandall-Liggett semigroup theory. The main difficult for this class of equations is due to the fact that (2) has a strongly degenerate term, i.e. when|Du|= 0. To overcome it, we employ the regularization method. We firstly prove the existence and uniqueness for weak solutions of the approximate p-Laplace equations. Then some prior estimate are given. Last by a standard limiting process as p→1+, we recover the well-posedness of the problem (1)-(6).
In the second part of this chapter, Analysising the models in [6] and [8], we combine the two models and propose the following model based on the reaction-diffusion partial differential system: where
Gσ(x) is the Gaussian kernel, namely
T> 0, A> 0,f is the original image,κ1and k2 are positive constants,Ωis a bounded domain of RN with the smooth boundary, and n denotes the unit outward normal to the boundary (?)Ω. In [8], we change (?)|▽υ|p(χ) by |▽υ|, and obtain the following model:
The corresponding Euler-Lagrange is as follows
Note that our model is the evolutionary form of the steady system (15). The main distinguish between the problem (7)-(12)and the problem (1)-(6) is the principal part of the diffusion equation onυ, and in the problem (7)-(12), the diffusion is p(χ)-Laplace. From (13) and (14), g(x) and p(x) depend on the location, x, in the original image. This way the speed and direction of diffusion at each location depends on the local behavior. The term|▽u| has larger value on the edges than in the homogeneous region, while g(x) is just the opposite. Therefore, the diffusion applied by equation (7) may cause little smoothing near the edges. The main feature of equation (7) is the variable character of nonlinearity which causes a gap between the homogeneous regions and nonhomogeneous regions (edges). So it ensures TV based diffusion (as p→1) along edges and Gaussian smoothing (as p→2) in homogeneous regions. On the other hand, as far as we know, there is few literature about PDE system devoted to image processing. Different from the models as before, in the new system, the two equations interact with each other:first we use (7) which is from the mixed diffusion model to obtain the smoother imageυ. then we obtainωfrom (8) to adjust the fidelity betweenυand f in (7). In other words, the denoising equation (7) has a different fidelity termωwhich is also derived from an evolutionary equation, thus we can select an appropriate scale to achieve an appropriate fidelity betweenυand f and obtain a better result for noise removal and edge or texture preservation. Furthermore, the new mode is to work directly on the equations, without thinking of any energy. Starting from the initial image f(χ) and by running the system (7)-(12) we construct a family of functions (i.e. the smoother images){υ(t,x)}t>o representing successive versions of f(x). we can choose the best image for the appropriate t and observe the whole evolution process. Finally, using the Galerkin method, we can proveυis the element of belong to Orlicz-Sobolev spaces W1,p(χ)(QT) (the rigorous definition is given in Section 2 below) theoretically, which allows for better regularity, and therefore achieve a good denoising result.
In the second chapter, we propose the following model where
where, Ga(x) is the Gaussian kernel,f is the original image,κ1> 0,κ2> 0,σ1,σ2 and T> 0 are fixed constants, c is a constant dependent ofκ1 andσ1,Ωis a bounded open domain of RN with the appropriate smooth boundary, n denotes the unit outward normal to the boundary (?)Ω. Let
where
It is clearly that the model can be rewrite as following
On the one side, we detailedly analysis some properties of the function C(s), and show that the edge detection function C(s) is likely to that of the original Perona-Malik diffusion, while the new diffusion equation has not backward diffusion. On the other side, letting g(s)= 1, as s→+∞, p(s)→1, the divergence principal part of the proposed diffusion equation is likely to the divergence operator of the mean curvature diffusion equation[13]; while, as s= 0, p(s)≈2, the divergence principal part is the diffusion term of the heat equation. Because of C(s), the new model has a hybrid diffusion type combined the mean curvature diffusion with the heat diffusion, and g(s) enhances the efficacy of C(s). In theory, using the fixed point theorem, we verify the existence and uniqueness of the solutions for the problem (16)-(18). And then we continue to verify the stability, boundedness, and the behavior of the solutions as t→∞, and these properties influence the results of image restoration. Next, we propose a convergent iterative scheme and verify the convergence of the scheme. Finally, experimental results illustrate the effectiveness of the new models with respect to other known models.
In the third chapter, using the CV model, we propose two fast image segmen-tation algorithm. The first model is primarily to improve the speed of level-sets segmentation methods. we develop a fast level set method for image segmentation without solving the Euler-Lagrange equation of the underlying variational problem proposed by Chan and Vese. Compared with the original CV model, our approach is a new practical fast way to solve the active contour without edge problem and avoids the process of complicated, expensive reinitialization. Since the calculation cost of the second step is O(N), even for the larger images, it is also very efficient. The second model is mainly to improve the color image segmentation results. From the idea of some the color image denoising models, we propose the color image seg-mentation model based on a convex energy functional, and using the principle of duality, we obtain the corresponding algorithm. From the numerical experiments, we can see the superiority of our model.
在第一章里,我们提出两个用于图像去噪和分解的反应扩散方程组模型.在第一部分中,受到Osher等人提出的基于H-1模的图像恢复和分解泛函模型的启发,我们提出一个耦合的反应扩散方程组模型.新模型包含两个主部不同的方程：其中一个主部为全变差扩散,用来去除噪声；另外一个主部为热扩散,用来修正去噪方程的源项,达到保护图像纹理的效果.两个方程相互作用,最终达到去噪的同时又能很好的保持边界和纹理.在第二部分中,在上述方程组模型的框架下,我们引入p(x)-Laplace流来代替原来的全变差流,从而新的去噪方程结合Gaussian热方程的各向同性扩散和全变差流的各向异性扩散的特点,达到自适应的效果.在理论研究方面,利用正则化方法以及Galerkin方法我们分别得到两个模型解的适定性,这为数值计算提供了必要的准备.与原来的模型不同,新模型基于反应扩散方程组框架.另外,新模型在去除阶梯效应、保持边界和去噪方面都有很好的效果,尤其在保持纹理方面效果显著.
在第二章里,我们提出用于图像去噪的混合型扩散方程模型.通过分析图像边界与图像位置和结构的依赖关系,构造出一个抗噪能力较强的边界映射函数,对应的扩散方程不仅具有各向异性扩散的特点,而且在图像内部齐次区域进行Gaussian热扩散,在近边界区域进行平均曲率扩散,最终达到自适应去噪的效果.在理论研究方面,我们首先利用不动点方法证明新模型的适定性,随后研究t→∞时解的渐近状态,结论表明新模型恢复结果的长时间极限状态为初始图像的局部均值.最后,我们提出一种计算格式,分析其收敛性,并对一系列图像进行数值实验.与TV模型相比,新模型在内部齐次区域进行Gaussian热扩散,有效避免阶梯效应；与PM模型相比,新模型是适定的,从而解更加稳定,而且由于新模型的扩散模式和扩散速度依赖于图像的位置和结构信息,模型自适应能力更强,在内部齐次区域能更好的去噪,在近边界处能更精细的保持边界.
在第三章中,我们基于活动轮廓方法建立两个图像分割模型.第一个模型主要用于加快基于活动轮廓的水平集分割算法.通过分析CV无边界活动轮廓水平集分割算法对光滑度不同的图像的分割结果,我们提出一种新的图像分割快速算法.新算法去掉原模型的正则项,在灰度水平集上离散泛函能量进行计算,计算复杂度为O(N),计算时间很短,而且分割结果和原始模型相近.第二个模型主要是改进彩色图像的分割效果.基于偏微分方程彩色图像去噪模型,我们提出一个凸的能量泛函向量值图像分割模型,并根据对偶原理,构造出求解新模型的算法.数值实验显示新算法的分割结果更加符合实际要求,而且新算法也相对快速稳定.
The image is between the humanity exchanges one of most direct-viewing ways. Especially in current information technology rapid development time, image infor-mation as an important communication medium are widely used in multimedia, digital health, artificial intelligence, aerospace and geological remote sensing, and other fields. Generally speaking, the digital image processing technology take the digital image as an object, including image acquisition, image analysis and image understanding three main areas. Image analysis is between image acquisition and image understanding between one of the links, it directly affects computer vision, robot vision, and machine vision. In image processing, image restoration and image segmentation is important problem and has been extensively studied, they all belong to the image analysis of research areas.
Image processing research has a long history, goes back to the 1920s, most of the image processing methods are based on a dimensional signal processing technologies, such as filtering technology, statistical theory, etc. With remote sensing, medical, and other fields growing demands, as well as physical technology breakthrough, peo-ple are increasingly concerned about the nature of image processing, and attempt to use strict mathematical theory to the current image processing method for colla-tion and improvement. Currently, the image processing technology has three main tools:Stochastic Models, Wavelet Analysis Theory and Partial Differential Equa-tions Methods. The first two have a long history:Stochastic Models are based on Bayes and markov random field and are to work directly on the images; Wavelet Analysis Theory is based on Fourier analysis, transform the image information to the frequency domain information, and then are used for image processing. Since the 1990s, image processing methods based on PDEs (partial differential equations) developed rather quickly, becoming a very attractive study areas. At present, in im-age restoration, image segmentation, image reconstruction, image recognition, image analysis and so on, image processing methods based on PDEs has been widely ap-plied. In PDEs models for image processing, image information is quantified as a continuous two-dimensional function, and can be seen as a field or physical state. PDEs methods work directly on the image, is closer to reality the objective world and easy to analysis models. In the mathematical theory, there are lots of the the-ories on PDEs, and then the models can be deeply analyzed. From this stage of development, we can not know which is the best in these three tools, And these three tools is kept in touch with each other and cross-promotion.
In this article, the first problem we are concerning is PDEs methods for image restoration and decomposition and the corresponding theoretical analysis. In image processing, image restoration has always been one of the most important fundamen-tal research subjects. In the image acquisition, copy, scan, transmit, display, and so on,the images will be inevitably degraded, such as blur and noise. However, in many application areas, people need a clear, high-quality images, so image restora-tion (such as denoising, deblurring, etc) is of great significance. In theory, noise is unpredictable random error, and subject to some probability distribution. Assume that u:Ω(?) R2→R is the intensity of the ideal image, n(x) is the noise,f is the
intensity of the observed image. The idea image u and the observed image f are
related by
f= u+n.
Generally speaking, the noise n(x) satisfies:
where, the first constrain indicates that the noise has zero mean, and the second one uses a priori information that the standard deviation of the noise n(x) isη. For ease of analysis, we assume that the image signal and noise are.independent of each other. The problem is to recover the restoration image, u, from the observed, noisy image, u0, where the two are related by u0= u+n(χ). The image restoration and decomposition models based on PDEs have the two important framework:the variational methods based on the optimal problem for energy functionals and the diffusion equation methods based on fluid diffusion theory. Image decomposition is the extension of image noise, Image denoising model basically can also be used as image decomposition model. Next, we present some image denoising models based on the variational method, image decomposition models based on the variational method, and image restoration models based on diffusion equations.
First, we introduce image denoising models based on the variational method and some classical denoising models. The famous classical and general method for removing Gaussian noise firs consists in solving the following constrained minimiza-tion problem:
λ> 0 is a weight parameter. The first term in E(u) is a smoothing term for deniosing. The second term in E(u) measures the fidelity to the initial data. In
1977, in the model proposed by Tikhonov and Arsenin [1],(?)(s)= s2. And the solution space is
whereΩis a bounded open domain of-RN with the appropriate smooth boundary, n denotes the unit outward normal to the boundary (?)Ω, and then the constrained minimization problem has unique solution in W. However, the above-mentioned (?) punish the gradient too much and W1,2(Ω) is a smoother space, and therefore the recovery image is too smooth to preserve the image edge information. The real image maybe not be smooth, so someone try to find a better function (?), for example, (?)(s)=|s|p,1< p< 2, and then the corresponding solution space is W1,p(Ω). In 1992,Rudin, Osher and Fatemi [2] propose the famous total variation method, the ROF model. In the TV model,(?)(s)=|s|, and mathematically this is reasonable, since it is natural to study solutions of this problem in the space of functions of bounded variation, BV(Ω), allowing for discontinuities which are necessary for edge reconstruction. The TV model has been studied extensively and has proved to be an invaluable tool for preserving edges in image restoration problem. In 1997, Chambolle and Lions [3] detailed theoretical analysis for the TV model, and propose the corresponding numerical calculation method. In 2003, Rudin, Lions and Osher [4] summary theoretical results for the total variation model, as well as the related theory of partial differential equations. When (?)(s)= (1+s2)~(1/2), the corresponding model is based on the minimal surface functional, and the functional has similar properties with the TV model and well-posedness [13]. When (?)(s) is the non-convex function, such as(?)(s)= log(1+s2)[9], using the model, we can have better results for image restoration. However,the existence and uniqueness of the solutions for the model still are a open problem, the convergence of the numerical scheme for the model had not been proved. The TV model is well posed, but TV-based denoising favors the piecewise constant solutions. Sometimes this also causes a staircasing effect in which noisy smooth regions are processed into piecewise constant regions [3], [5], and [38]. Not only'blocky'solutions fail to satisfy the ubiquitous'eyeball norm' but they can also develop'false edges' which can mislead a human or computer into identifying erroneous features not present in the true image. Some authors consider another regularizing term to remove the noise [5], which is as follows
where (?) p(s)→2, (?) p(s)→1, and p is monotonically decreasing. This model should reap the benefits of both isotropic and TV-based diffusion, as well as a com-bination of the two. However, it is difficult to study mathematically since the lower semi-continuity of the functional is not readily evident. In [6], Chen, Levine, and Rao modify the model, and propose a functional with variable exponent,1≤p(x)≤2, which is a combination of Total Variation based regularization and Gaussian smooth-ing.
Secondly, we introduce some image decomposition model based on the varia-tional method. If we decompose the image into two components:the smoothed version of the original image f, denoted by u, which represents the cartoon in-formation of the original image, and the componentυ= f - u, which represents the texture or noise information, then the above model is actually a new energy functional for image restoration and image decomposition into cartoon plus texture, where the first term is a regularizing term to remove the noise, and the second term is a fidelity term. Though the ROF model performs very well for image denoising and edge protecting, it may also destroy small details, such as textures, see [7]. To overcome this drawback, Meyer [7] proposed a new minimization method by intro-ducing a weaker norm which is more appropriate to represent textured or oscillatory patterns. Subsequently, Osher et al. [8] presented the following model which com- bined the norm for oscillatory functions proposed by Meyer [7] (involving the H -1 norm) with the total variation minimization from the ROF model [2] whereλ> 0 is a weight parameter. Osher, Sole and Vese[8] is in turn a particular case of an earlier work by Osher and Vese [24]. Inspired by Meyer's G-norm, some authors have proposed various numerical models for image decomposition and mod-eling texture with Meyer's G-norm and F-norm, and also with Besov norms[25]-[27].
Finally, we present and analysis some denoising models based on the diffusion equations noise model, which is from the fluid diffusion physical phenomena. Early in the 1960s, the oldest and most investigated equation proposed by Gabor in the processing is probably the parabolic linear heat equation with the initial data f. Alvarez et al. [10] rigorously establish the connection between scale space analysis and PDEs. In [9], Perona and Malik developed an anisotropic diffusion scheme for image denosing. The basic idea of this nonlinear smoothing scheme was to smooth the image while preserving the edges in it. This was done by using the equation
where C(s) is a smooth non-increasing function such that C(s)→1, as s→0 and C(s)→0, as s→∞. In fact, the PM diffusion equation is a forward-backward diffusion and the well-posedness of the equation is still a open problem. It is worth mentioning that, the TV model can be seen as the Perona-Malik diffusion when c(|▽υ|)= (?). In [11], A simple adjustment with practical applications, is to include a short range mollifier in the nonlinear diffusion. Existence and uniqueness of solutions to this modified Perona-Malik equation has been proved by Catte et al. for initial dataυ0∈L2(Ω). Inspired by the PM model, many researchers proposed new models which is to work on the nonlinear anisotropic diffusion equations, without thinking of any energy functional. All models are nonlinear anisotropic diffusion filters with diffusivities being adapted to the local image structure. Weickert [12] propose the model which is a nonlinear anisotropic diffusion filter with diffusion tensors being adapted to the local image structure. The author modify the diffusion tensor matrix eigenvectors and eigenvalues to control the diffusion direction and diffusion strength. The models based on PDEs can easily extend to the color image case. In the color image case, the geometry properties of surfaces [13] is involved. In addition, the convection term in PDEs is introduced, and can speed up calculation speed [14]-[15]. In practice, the ill-posedness results in a mild instability in the PM problem. Regions of high gradients develop a "staircase" instability that involves dynamic coarsening of the steps as time evolves. To make the images more pleasing to the eye, it would be useful to reduce this effect. To this end, many researchers consider higher-order version of the Perona-Malik equations, examples of which.are given in [16]-[18].
In this article, the second problem we are concerning is image segmentation models based on PDEs and the level-sets method. Image segmentation is one of the most fundamental problems in the fields of image processing and computer vision. The aim is to find a partition of an image into its constituent parts. However, this definition is rather unsatisfactory and ambiguous. As we will see, the main difficulty is that one need to manipulate objects of different kinds:functions, domains in R2, and curves. In the last century in the late 1980s, partial differential equation method is first applied to the image segmentation. Kass, Witkin and Terzopoulos[19] first propose the snake and the geodesic active contours models. Mumford and Shah [20] propose the variational segmentation method for the region detection. Since the image edges is one-dimensional, in early segmentation algorithms it is difficult to indicate the edge information for images. At that time, using theΓ-convergence or spline curves and other techniques, the researchers obtain the image segmentation results. However, these methods are not only difficult to understand, and solve complex, and do not have the expected effect.1988年, Osher and Sethian [21] introduce the level-sets method which allows for cusps, corners, and automatic topological changes. The curve is represented implicitly via a Lipschitz function, by the zero level-set of the Lipschitz function, and the evolution of the curve is given by the zero-level curve from the evolution of the function. Since then, the PDEs method is widely used to image segmentation. The iconic work can be see the geodesic active contours models by Caselles, Kimmel and Sapiro[22] and the active contours without edges models by Chan and Vese [23]. At present, the CV model has many improvements and development and has attracted the attention of many scholars.
In this paper we studies image processing methods based on PDEs, especially in image restoration, decomposition and segmentation applications. Since in im-age processing the models based on PDEs (system) which is from the research and analysis of the image structure, is usually the new type diffusion equation (system), so the existence and uniqueness of solutions for the new PDEs(system) is impor-tant practical significance for the numerical calculation. In this paper the following problems are studies:the first chapter is the theoretical research and numerical implementation for the image restoration and decomposition model based on new partial differential systems; the second chapter is the theoretical research and nu-merical implementation for the image restoration model based on new anisotropic diffusion equations; the third chapter is the introduction and analysis of two new fast segmentation algorithms. every problem in this paper can be divided into the following sections:first, we introduce a new model, usually based on some diffu-sion equation (system); then, we have demonstrated the existence, uniqueness and boundedness of solutions for the new models; finally, we structure the numerical scheme which is applied to image processing with respect to other known models. Since the new equation (system) structure is different, we applied the regularization method, the Galerkin method, and the fixed-point method to obtain the existence of solutions for the new models, respectively. Then, experimental results illustrate the effectiveness of the new models with respect to other known models.
In the first chapter, the two image restoration and decomposition models based on reaction-diffusion PDEs (system) is proposed. In the first part of this chapter, inspired by Osher, Sole and Vese[8], we propose the following model based on the reaction-diffusion partial differential system: where f is the original image,Ωis a bounded domain of RN with appropriately smooth boundary, n the unit outward normal toΩ,1≤N< 4, T> 0,λ> 0, and as far as we know, there is few literature about PDE system devoted to image processing. Different from the models as before, in the new system, the two equations interact with each other:first we obtainωfrom (1) to adjust the fidelity betweenυand f in (2), then we use (2) which is from the classical TV model to obtain the smoother image u. In other words, the denoising equation (2) has a different fidelity termωwhich is also derived from an evolutionary equation, thus we can select an appropriate scale to achieve an appropriate fidelity betweenυand f and obtain a better result for noise removal and edge or texture preservation. Furthermore, we can proveυis the element of BV space theoretically, which allows for discontinuities along curves, therefore edges and contours are kept in the image u. In addition, in theory, the problem (2) can be seen as a TV flow equation with nonlinear source. In [28,29,30], Andreu et al. studied the TV flow ut= div(Du/|Du|) with L1-data using the Crandall-Liggett semigroup theory. The main difficult for this class of equations is due to the fact that (2) has a strongly degenerate term, i.e. when|Du|= 0. To overcome it, we employ the regularization method. We firstly prove the existence and uniqueness for weak solutions of the approximate p-Laplace equations. Then some prior estimate are given. Last by a standard limiting process as p→1+, we recover the well-posedness of the problem (1)-(6).
In the second part of this chapter, Analysising the models in [6] and [8], we combine the two models and propose the following model based on the reaction-diffusion partial differential system: where
Gσ(x) is the Gaussian kernel, namely
T> 0, A> 0,f is the original image,κ1and k2 are positive constants,Ωis a bounded domain of RN with the smooth boundary, and n denotes the unit outward normal to the boundary (?)Ω. In [8], we change (?)|▽υ|p(χ) by |▽υ|, and obtain the following model:
The corresponding Euler-Lagrange is as follows
Note that our model is the evolutionary form of the steady system (15). The main distinguish between the problem (7)-(12)and the problem (1)-(6) is the principal part of the diffusion equation onυ, and in the problem (7)-(12), the diffusion is p(χ)-Laplace. From (13) and (14), g(x) and p(x) depend on the location, x, in the original image. This way the speed and direction of diffusion at each location depends on the local behavior. The term|▽u| has larger value on the edges than in the homogeneous region, while g(x) is just the opposite. Therefore, the diffusion applied by equation (7) may cause little smoothing near the edges. The main feature of equation (7) is the variable character of nonlinearity which causes a gap between the homogeneous regions and nonhomogeneous regions (edges). So it ensures TV based diffusion (as p→1) along edges and Gaussian smoothing (as p→2) in homogeneous regions. On the other hand, as far as we know, there is few literature about PDE system devoted to image processing. Different from the models as before, in the new system, the two equations interact with each other:first we use (7) which is from the mixed diffusion model to obtain the smoother imageυ. then we obtainωfrom (8) to adjust the fidelity betweenυand f in (7). In other words, the denoising equation (7) has a different fidelity termωwhich is also derived from an evolutionary equation, thus we can select an appropriate scale to achieve an appropriate fidelity betweenυand f and obtain a better result for noise removal and edge or texture preservation. Furthermore, the new mode is to work directly on the equations, without thinking of any energy. Starting from the initial image f(χ) and by running the system (7)-(12) we construct a family of functions (i.e. the smoother images){υ(t,x)}t>o representing successive versions of f(x). we can choose the best image for the appropriate t and observe the whole evolution process. Finally, using the Galerkin method, we can proveυis the element of belong to Orlicz-Sobolev spaces W1,p(χ)(QT) (the rigorous definition is given in Section 2 below) theoretically, which allows for better regularity, and therefore achieve a good denoising result.
In the second chapter, we propose the following model where
where, Ga(x) is the Gaussian kernel,f is the original image,κ1> 0,κ2> 0,σ1,σ2 and T> 0 are fixed constants, c is a constant dependent ofκ1 andσ1,Ωis a bounded open domain of RN with the appropriate smooth boundary, n denotes the unit outward normal to the boundary (?)Ω. Let
where
It is clearly that the model can be rewrite as following
On the one side, we detailedly analysis some properties of the function C(s), and show that the edge detection function C(s) is likely to that of the original Perona-Malik diffusion, while the new diffusion equation has not backward diffusion. On the other side, letting g(s)= 1, as s→+∞, p(s)→1, the divergence principal part of the proposed diffusion equation is likely to the divergence operator of the mean curvature diffusion equation[13]; while, as s= 0, p(s)≈2, the divergence principal part is the diffusion term of the heat equation. Because of C(s), the new model has a hybrid diffusion type combined the mean curvature diffusion with the heat diffusion, and g(s) enhances the efficacy of C(s). In theory, using the fixed point theorem, we verify the existence and uniqueness of the solutions for the problem (16)-(18). And then we continue to verify the stability, boundedness, and the behavior of the solutions as t→∞, and these properties influence the results of image restoration. Next, we propose a convergent iterative scheme and verify the convergence of the scheme. Finally, experimental results illustrate the effectiveness of the new models with respect to other known models.
In the third chapter, using the CV model, we propose two fast image segmen-tation algorithm. The first model is primarily to improve the speed of level-sets segmentation methods. we develop a fast level set method for image segmentation without solving the Euler-Lagrange equation of the underlying variational problem proposed by Chan and Vese. Compared with the original CV model, our approach is a new practical fast way to solve the active contour without edge problem and avoids the process of complicated, expensive reinitialization. Since the calculation cost of the second step is O(N), even for the larger images, it is also very efficient. The second model is mainly to improve the color image segmentation results. From the idea of some the color image denoising models, we propose the color image seg-mentation model based on a convex energy functional, and using the principle of duality, we obtain the corresponding algorithm. From the numerical experiments, we can see the superiority of our model.
引文
[1]A. N. Tikhonov and V. Y.'Arsenin, Solutions of ill-posed problems, Winston and Sons, Washington, D.C.,1977.
[2]L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algo-rithms, Physica D 60 (1992) 259-268.
[3]A. Chambolle and P.L. Lions, Image recovery via total variation minimization and related problems, Numer. Math.,76 (1997), pp.167-188.
[4]L. Rudin, P. L. Lions and S. Osher, Multiplicative denoising and deblurring:Theory and algorithms, Geometric Level Set Methods in Imaging, Vision and Graphics, S. Other and N. Paragios, eds., New York,2003,103-119.
[5]P. Blomgren, T. F. Chan, P. Mulet, and C. Wong, Total variation image restoration: Numerical methods and extensions, Proceedings of the 1997 IEEE International Conference on Image Processing, III (1997), pp.384-387.
[6]Yunmei Chen, Stacey Levine, and Murali Rao, Variable Exponent, Linear Growth Functionals in Image Restoration, SIAM J. Appl. Math.4(66)(2006),1383-1406.
[7]Y. Meyer, Oscillating Patterns in Image Processing and Nonlinear Evolution Equa-tions, Univ. Lecture Ser.22, AMS, Providence, RI,2002.
[8]S. Osher, A. Sole, L. Vese, Image decomposition and restoration using total varia-tion minimization and the H-1 norm, SIAM Journal on Multiscale Modeling and Simulation,1(3) (2003),349-370.
[9]P. Perona, J. Malik, Scale space and edge detection using anisotropic diffusion, IEEE Trans. Pattern Anal. Mach. Intell.12 (1990),629-639.
[10]L. Alvarez, F. Guichard, P. L. Lions, and J. M. Morel, Axioms and fundamental equation of image processing, Archive for Rational Mechanics and Analysis,123 (3)(1993),199-257.
[11]F. Catte, P.L. Lions, J.M. Morel, and T. Coll, Image selective smoothing and edge detection by nonlinear diffusion, SIAM J. Numer. Anal.129 (1992),182-193.
[12]J. Weickert, Theoretical foundations of anisotropic diffusion in image processing, Computing Supplement,11 (1996),221-236.
[13]A. I. El-Fallah and G. E. Ford, On mean curvature diffusion in nonlinear image filtering, Pattern Recognition Letters,19(1998),433-427.
[14]Y. Chen, B. C. Vemuri, L. Wang, Image denoising and segmentation via nonlinear diffusion, Computers and Mathematics with Applications,39(2000),131-149.
[15]C. A. Z. Barcelos, and Y. Chen, Heat flows and related minimization problem in image restoration, Computers and Mathematics with Applications,39(2000),81-97.
[16]Jack Tumblin and Greg Turk. LCIS:A boundary hierarchy for detail-preserving contrast reduction. In Proceedings of the SIGGRAPH 1999 annual conference on Computer graphics, August 8-13,1999, Los Angeles, CA USA, pages 83-90,1999.
[17]Guo W. Wei. Generalized Perona-Malik equation for image processing. IEEE Signal Processing Letters,6(7):165-167, July 1999.
[18]Yu-Li You and M. Kaveh. Fourth-order partial differential equations for noise re-moval. IEEE Trans. Image Process.,9(10):1723-1730,2000.
[19]M. Kass, A. Witkin and D. Terzopoulos, Snakes:Active contour models, In First International Conference on Computer Vision, (1987),259-268.
[20]D. Mumford and J. Shah, Boundary detection by minimizing functionals, In Proceed-ings of the International Conference on Computer Vision and Pattern Recognition, San Francisco, CA,1985,22-26.
[21]S. Osher and J. Sethian, Fronts propagating with curvature dependent speed:al-gorithms based on the Hamilton-Jacobi formulation, Journal of Computational Physics,79(1998),12-49
[22]Caselles, V., Kimmel, R., Sapiro, G. Geodesic active contours.Int. J. Comput. Vis., 22(1)(1997):61-79.
[23]T. F. Chan and L. Vese, Active contours without edges, IEEE Transactions on Image Processing,10(2001),266-277.
[24]L. Vese and S. Osher, Modeling textures with total variation minimization and oscillating patterns in image processing, J. Sci. Comput.,19(1-3) (2003),553-571.
[25]J. Garnett, T. Le, L. Vese, Image decomposition using bounded variation and gen-eralized Besov spaces, UCLA CAM Report 05-57, October 2005.
[26]T. Le, L. Vese, Image decomposition using total variation and div(BMO), Multiscale Model. Simul.,4(2)(2005),390-423.
[27]L. Lieu, L. Vese, Image restoration and decomposition via bounded total variation and Negative Hilbert-Sobolev spaces, UCLA CAM Report 05-33, May 2005.
[28]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, Minimizing total variation flow, C. R. Acad. Sci. Paris Ser. I Math.,331(2000), no.11,867-872.
[29]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, The Dirichlet problem for the total variation flow, J. Funct. Anal.,180(2001), no.2,347-403.
[30]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, Minimizing total variation flow, Differential Integral Equations,14(2001), no.3,321-360.
[31]L. C. Evans and R. F. Gariepy,Measure Theory and Fine Properties of Functions, Stud. Adv. Math., CRC Press, Boca Raton, FL,1992.
[32]O. A. Ladyzhenskaja, V. A. Solonnikov and N. N. Ural'tzeva, Linear and quasilinear equations of parablolic type, Transl. Math. Mono.,23, AMS. Providence RI,1968.
[33]R. Acar and C. R. Vogel, Analysis of bounded variation penalty methods for ill-posed problems, Inverse Problems,10 (1994), pp.1217-1229.
[34]V. Caselles, J.-M. Morel, and C. Sbert, An axiomatic approach to image interpola-tion, IEEE Trans. Image Process.,7 (1998), pp.376-386.
[35]D. M. Strongand T. F. Chan, Spatially and Scale Adaptive Total Variation Based Regularization and Anisotropic Diffusion in Image Processing, Technical report, University of California, Los Angeles, CA, CAM96-46(UCLA).
[36]L. Vese and S. Osher, Modeling textures with total variation minimization and oscillating patterns in image processing, Journal of Scientific Computing,19(1-3) (2003),553-572.
[37]C. M. Elliott and S. A. McBeth, Analysis of the TV Regularization and H-1 Fidelity Model for Decomposing an Image into Cartoon plus Texture, Communications Pure and Applied Analysis,6(4) (2007),917-936
[38]M. Nikolova, Weakly constrained minimization:application to the estimation of images and signals involving constant regions, J. Math. Imaging Vision,21(2004), 155-175.
[39]D. C. Dobson and F. Santosa, Recovery of blocky images from noisy and blurred data, SIAM J. Appl. Math.,56(1996),1181-1198.
[40]W. Ring, Structural properties of solutions to total variation regularization prob-lems, M2AN Math. Model. Numer. Anal.,34(2000),799-810.
[41]R. T. Whitaker and S. M. Pizer, A mulit-scale approach to nonuniform diffusion, Comput. Vis. Graph. Image Process.:Image Understand.,57(1993),99-110.
[42]Y.-L. You, W. Xu, A. Tannenbaum, and M. Kaveh, Behavioral analysis of anisotropic diffusion in image processing, IEEE Trans. Image Process.,5(1996), 1539-1553.
[43]S. Antontsev and S. Shmarev, Extinction of solutions of parabolic equations with variable anisotropic nonlinearities, Proceedings of the Steklov Institute of Mathemat-ics,261(1)(2008),11-21.
[44]Z.Q. Wu, Q.N. Zhao, J.X. Yin and H.L. Li, Nonlinear Diffusion Equations, World Scientific, Singapore,2001.
[45]L. Diening, Maximal function on generalized Lebesgue spaces LP(·),Math. Inequal. Appl.,7(2004),245-253.
[46]D. Edmunds, J. Rakosnik, Sobolev embeddings with variable exponent, Studia Math.,143(3)(2000),267-293.
[47]O. Kovacik, J. Rakosnik, On spaces Lp(χ) and Wκ,p(χ)Czechoslovak Math. J. 41(116)(1991),592-618.
[48]J. Musielak, Orlicz spaces and modular spaces, vol.1034 of Lecture Notes in Math-ematics, Springer-Verlag, Berlin,1983.
[49]V. V. Zhikov, On the density of smooth functions in Sobolev-Orlicz spaces, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI),310 (2004),67-81, 226.
[50]P. Harjulehto, P. Hasto, An overview of variable exponent Lebesgue and Sobolev spaces, in Future trends in geometric function theory, vol.92 of Rep. Univ. Jyvaskyla Dep. Math. Stat., Univ. Jyvaskyla, Jyvaskyla,2003,85-93.
[51]S. Samko, On a progress in the theory of Lebesgue spaces with variable exponent: maximal and singular operators, Integral Transforms Spec. Funct.,16 (2005), pp. 461-482.
[52]A. Kufner, J. Oldrich, and S. Fucik, Function Spaces, Kluwer Academic Publishers, 1977.
[53]J. Simon, Compact sets in the space Lp(0, T;B), Ann. Mat. Pura Appl., IV. Ser. 146 (1987),65-96.
[54]Tai, Xue-Cheng; Lie, Knut-Andreas; Chan, Tony F.; Osher, Stanley (eds):Image Processing based on Partial differential equations, "Mathematics and Visualiza-tion", Heidelberg:Springer Verlag 2006. ISBN 978-3-540-33266-4.440 pages.
[55]Qiang Liu, Zhengan Yao and Yuanyuan Ke, Entropy solutions for a fourth-order nonlinear degenerate problem for noise removal. Nonlinear analysis, TMA, 67(2007),1908-1918.
[56]R. T. Whitaker and S. M. Pizer, A mulit-scale approach to nonuniform diffusion, Comput. Vis. Graph. Image Process.:Image Understand.,57 (1993),99-110.
[57]M. Nitzberg and T. Shiota, Nonlinear image filtering with edge and corner en-hancement, IEEE Transactions on Pattern Analysis and Machine Intelligence 14 (8),826-832, (1992).
[58]L. Alvarez, P.L. Lions and J.M. Morel, Image selective smoothing and edge detection by nonlinear diffusion. II, SIAM J. Numer. Anal.29 (3),845-866, (June 1992).
[59]Vicent Caselles, Jean-Michel Morel, Guillermo Sapiro, and Allen Tannenbaum. In-troduction to the special issue on partial differential equations and geometry-driven diffusion in image processing and analysis. IEEE Transactions on image processing, 7(3):269-273, March 1998.
[60]Selim Esedoglu. An analysis of the Perona-Malik Scheme. PhD thesis, Courant Institute of Mathematical Sciences, NYU, July 2000.
[61]R. Whitaker and S. Pizer. A multi-scale approach to nonuniform diffusion. CVGIP:Image Understanding,57(1):99-110,1993.
[62]G. Aubert and P. Kornprobst. Mathematical Problems in Image Processing:PDE's and the Calculus of Variations, vol.147 of App. Mathem. Sciences. Springer-Verlag, January 2002.
[63]L. C. Evans, Partial differential equations, Graduate Studies in Mathematics,19, American Mathematical Society, Providence, Rhode Island,1998.
[64]R. Adams, Sobolev spaces, in:Pure and Applied Mathematics Series of Monographs and Textbooks, vol.65, Academic Press, Inc., New York, SanFrancisco, London, 1975.
[65]S. Osher and N. Paragios. Geometric Level Set Methods in Imaging, Vision, and Graphics. Springer-Verlag,2003. New York, NY.
[66]Kimmel R.,Fast edge integration, In Level Set Methods and Their Applications in Computer Vision, Chapter 3. Springer Verlag:NY.2003.
[67]A. Tsai, A. Yezzi, and A. S. Willsky. Curve evolution implementation of the Mumford-Shah functional for image segmentation, denoising, interpolation, and magnification, IEEE Tran. Image Process.,10(8)(2001),1169-1186.
[68]Tony F. Chan, B. Yezrielev Sandberg, and Luminita A. Vese, Active Contours without Edges for Vector-Valued Images, Journal of Visual Communication and Image Representation,2(11)(2000),130-141.
[69]Weickert J., Romeny B. and Viergever M., Efficient and Reliable Scheme for Nonlinear Diffusion Filtering,IEEE Transactions on Image Processing,7 (3) (1998), 398-410.
[70]J. Weickert, Applications of nonlinear diffusion in image processing and computer vision. Acta Math. Univ. Comenian. (N.S.),70(1)(2000),33-50.
[71]Bresson X, Esedoglu S, Vandergheynst P, et al. Fast Global Minimization of the Active Contour/Snake Model. J Math Imaging Vis,28(2007),151-167.
[72]Guillermo Sapiro, Vector-Valued Active Contours, IEEE Computer Society Confer-ence on Computer Vision and Pattern Recognition (CVPR'96), (1996):680-685.
[73]Aujol J.-F., Gilboa G., Chan T., Osher S.:Structure-texture image decomposition-modeling, algorithms, and parameter selection. Int. J. Comput. Vis.67(1),111-136 (2006)Moggi E. Notions of computation and monads. Inform Comput,1991,93:55-92.
[74]Chan T., Esedoglu S., Nikolova M., Algorithms for finding global minimizers of image segmentation and denoising models, UCLA CAM Report 04-54 (2004).
[2]L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algo-rithms, Physica D 60 (1992) 259-268.
[3]A. Chambolle and P.L. Lions, Image recovery via total variation minimization and related problems, Numer. Math.,76 (1997), pp.167-188.
[4]L. Rudin, P. L. Lions and S. Osher, Multiplicative denoising and deblurring:Theory and algorithms, Geometric Level Set Methods in Imaging, Vision and Graphics, S. Other and N. Paragios, eds., New York,2003,103-119.
[5]P. Blomgren, T. F. Chan, P. Mulet, and C. Wong, Total variation image restoration: Numerical methods and extensions, Proceedings of the 1997 IEEE International Conference on Image Processing, III (1997), pp.384-387.
[6]Yunmei Chen, Stacey Levine, and Murali Rao, Variable Exponent, Linear Growth Functionals in Image Restoration, SIAM J. Appl. Math.4(66)(2006),1383-1406.
[7]Y. Meyer, Oscillating Patterns in Image Processing and Nonlinear Evolution Equa-tions, Univ. Lecture Ser.22, AMS, Providence, RI,2002.
[8]S. Osher, A. Sole, L. Vese, Image decomposition and restoration using total varia-tion minimization and the H-1 norm, SIAM Journal on Multiscale Modeling and Simulation,1(3) (2003),349-370.
[9]P. Perona, J. Malik, Scale space and edge detection using anisotropic diffusion, IEEE Trans. Pattern Anal. Mach. Intell.12 (1990),629-639.
[10]L. Alvarez, F. Guichard, P. L. Lions, and J. M. Morel, Axioms and fundamental equation of image processing, Archive for Rational Mechanics and Analysis,123 (3)(1993),199-257.
[11]F. Catte, P.L. Lions, J.M. Morel, and T. Coll, Image selective smoothing and edge detection by nonlinear diffusion, SIAM J. Numer. Anal.129 (1992),182-193.
[12]J. Weickert, Theoretical foundations of anisotropic diffusion in image processing, Computing Supplement,11 (1996),221-236.
[13]A. I. El-Fallah and G. E. Ford, On mean curvature diffusion in nonlinear image filtering, Pattern Recognition Letters,19(1998),433-427.
[14]Y. Chen, B. C. Vemuri, L. Wang, Image denoising and segmentation via nonlinear diffusion, Computers and Mathematics with Applications,39(2000),131-149.
[15]C. A. Z. Barcelos, and Y. Chen, Heat flows and related minimization problem in image restoration, Computers and Mathematics with Applications,39(2000),81-97.
[16]Jack Tumblin and Greg Turk. LCIS:A boundary hierarchy for detail-preserving contrast reduction. In Proceedings of the SIGGRAPH 1999 annual conference on Computer graphics, August 8-13,1999, Los Angeles, CA USA, pages 83-90,1999.
[17]Guo W. Wei. Generalized Perona-Malik equation for image processing. IEEE Signal Processing Letters,6(7):165-167, July 1999.
[18]Yu-Li You and M. Kaveh. Fourth-order partial differential equations for noise re-moval. IEEE Trans. Image Process.,9(10):1723-1730,2000.
[19]M. Kass, A. Witkin and D. Terzopoulos, Snakes:Active contour models, In First International Conference on Computer Vision, (1987),259-268.
[20]D. Mumford and J. Shah, Boundary detection by minimizing functionals, In Proceed-ings of the International Conference on Computer Vision and Pattern Recognition, San Francisco, CA,1985,22-26.
[21]S. Osher and J. Sethian, Fronts propagating with curvature dependent speed:al-gorithms based on the Hamilton-Jacobi formulation, Journal of Computational Physics,79(1998),12-49
[22]Caselles, V., Kimmel, R., Sapiro, G. Geodesic active contours.Int. J. Comput. Vis., 22(1)(1997):61-79.
[23]T. F. Chan and L. Vese, Active contours without edges, IEEE Transactions on Image Processing,10(2001),266-277.
[24]L. Vese and S. Osher, Modeling textures with total variation minimization and oscillating patterns in image processing, J. Sci. Comput.,19(1-3) (2003),553-571.
[25]J. Garnett, T. Le, L. Vese, Image decomposition using bounded variation and gen-eralized Besov spaces, UCLA CAM Report 05-57, October 2005.
[26]T. Le, L. Vese, Image decomposition using total variation and div(BMO), Multiscale Model. Simul.,4(2)(2005),390-423.
[27]L. Lieu, L. Vese, Image restoration and decomposition via bounded total variation and Negative Hilbert-Sobolev spaces, UCLA CAM Report 05-33, May 2005.
[28]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, Minimizing total variation flow, C. R. Acad. Sci. Paris Ser. I Math.,331(2000), no.11,867-872.
[29]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, The Dirichlet problem for the total variation flow, J. Funct. Anal.,180(2001), no.2,347-403.
[30]F. Andreu, C. Ballester, V. Caselles and J. M. Mazon, Minimizing total variation flow, Differential Integral Equations,14(2001), no.3,321-360.
[31]L. C. Evans and R. F. Gariepy,Measure Theory and Fine Properties of Functions, Stud. Adv. Math., CRC Press, Boca Raton, FL,1992.
[32]O. A. Ladyzhenskaja, V. A. Solonnikov and N. N. Ural'tzeva, Linear and quasilinear equations of parablolic type, Transl. Math. Mono.,23, AMS. Providence RI,1968.
[33]R. Acar and C. R. Vogel, Analysis of bounded variation penalty methods for ill-posed problems, Inverse Problems,10 (1994), pp.1217-1229.
[34]V. Caselles, J.-M. Morel, and C. Sbert, An axiomatic approach to image interpola-tion, IEEE Trans. Image Process.,7 (1998), pp.376-386.
[35]D. M. Strongand T. F. Chan, Spatially and Scale Adaptive Total Variation Based Regularization and Anisotropic Diffusion in Image Processing, Technical report, University of California, Los Angeles, CA, CAM96-46(UCLA).
[36]L. Vese and S. Osher, Modeling textures with total variation minimization and oscillating patterns in image processing, Journal of Scientific Computing,19(1-3) (2003),553-572.
[37]C. M. Elliott and S. A. McBeth, Analysis of the TV Regularization and H-1 Fidelity Model for Decomposing an Image into Cartoon plus Texture, Communications Pure and Applied Analysis,6(4) (2007),917-936
[38]M. Nikolova, Weakly constrained minimization:application to the estimation of images and signals involving constant regions, J. Math. Imaging Vision,21(2004), 155-175.
[39]D. C. Dobson and F. Santosa, Recovery of blocky images from noisy and blurred data, SIAM J. Appl. Math.,56(1996),1181-1198.
[40]W. Ring, Structural properties of solutions to total variation regularization prob-lems, M2AN Math. Model. Numer. Anal.,34(2000),799-810.
[41]R. T. Whitaker and S. M. Pizer, A mulit-scale approach to nonuniform diffusion, Comput. Vis. Graph. Image Process.:Image Understand.,57(1993),99-110.
[42]Y.-L. You, W. Xu, A. Tannenbaum, and M. Kaveh, Behavioral analysis of anisotropic diffusion in image processing, IEEE Trans. Image Process.,5(1996), 1539-1553.
[43]S. Antontsev and S. Shmarev, Extinction of solutions of parabolic equations with variable anisotropic nonlinearities, Proceedings of the Steklov Institute of Mathemat-ics,261(1)(2008),11-21.
[44]Z.Q. Wu, Q.N. Zhao, J.X. Yin and H.L. Li, Nonlinear Diffusion Equations, World Scientific, Singapore,2001.
[45]L. Diening, Maximal function on generalized Lebesgue spaces LP(·),Math. Inequal. Appl.,7(2004),245-253.
[46]D. Edmunds, J. Rakosnik, Sobolev embeddings with variable exponent, Studia Math.,143(3)(2000),267-293.
[47]O. Kovacik, J. Rakosnik, On spaces Lp(χ) and Wκ,p(χ)Czechoslovak Math. J. 41(116)(1991),592-618.
[48]J. Musielak, Orlicz spaces and modular spaces, vol.1034 of Lecture Notes in Math-ematics, Springer-Verlag, Berlin,1983.
[49]V. V. Zhikov, On the density of smooth functions in Sobolev-Orlicz spaces, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI),310 (2004),67-81, 226.
[50]P. Harjulehto, P. Hasto, An overview of variable exponent Lebesgue and Sobolev spaces, in Future trends in geometric function theory, vol.92 of Rep. Univ. Jyvaskyla Dep. Math. Stat., Univ. Jyvaskyla, Jyvaskyla,2003,85-93.
[51]S. Samko, On a progress in the theory of Lebesgue spaces with variable exponent: maximal and singular operators, Integral Transforms Spec. Funct.,16 (2005), pp. 461-482.
[52]A. Kufner, J. Oldrich, and S. Fucik, Function Spaces, Kluwer Academic Publishers, 1977.
[53]J. Simon, Compact sets in the space Lp(0, T;B), Ann. Mat. Pura Appl., IV. Ser. 146 (1987),65-96.
[54]Tai, Xue-Cheng; Lie, Knut-Andreas; Chan, Tony F.; Osher, Stanley (eds):Image Processing based on Partial differential equations, "Mathematics and Visualiza-tion", Heidelberg:Springer Verlag 2006. ISBN 978-3-540-33266-4.440 pages.
[55]Qiang Liu, Zhengan Yao and Yuanyuan Ke, Entropy solutions for a fourth-order nonlinear degenerate problem for noise removal. Nonlinear analysis, TMA, 67(2007),1908-1918.
[56]R. T. Whitaker and S. M. Pizer, A mulit-scale approach to nonuniform diffusion, Comput. Vis. Graph. Image Process.:Image Understand.,57 (1993),99-110.
[57]M. Nitzberg and T. Shiota, Nonlinear image filtering with edge and corner en-hancement, IEEE Transactions on Pattern Analysis and Machine Intelligence 14 (8),826-832, (1992).
[58]L. Alvarez, P.L. Lions and J.M. Morel, Image selective smoothing and edge detection by nonlinear diffusion. II, SIAM J. Numer. Anal.29 (3),845-866, (June 1992).
[59]Vicent Caselles, Jean-Michel Morel, Guillermo Sapiro, and Allen Tannenbaum. In-troduction to the special issue on partial differential equations and geometry-driven diffusion in image processing and analysis. IEEE Transactions on image processing, 7(3):269-273, March 1998.
[60]Selim Esedoglu. An analysis of the Perona-Malik Scheme. PhD thesis, Courant Institute of Mathematical Sciences, NYU, July 2000.
[61]R. Whitaker and S. Pizer. A multi-scale approach to nonuniform diffusion. CVGIP:Image Understanding,57(1):99-110,1993.
[62]G. Aubert and P. Kornprobst. Mathematical Problems in Image Processing:PDE's and the Calculus of Variations, vol.147 of App. Mathem. Sciences. Springer-Verlag, January 2002.
[63]L. C. Evans, Partial differential equations, Graduate Studies in Mathematics,19, American Mathematical Society, Providence, Rhode Island,1998.
[64]R. Adams, Sobolev spaces, in:Pure and Applied Mathematics Series of Monographs and Textbooks, vol.65, Academic Press, Inc., New York, SanFrancisco, London, 1975.
[65]S. Osher and N. Paragios. Geometric Level Set Methods in Imaging, Vision, and Graphics. Springer-Verlag,2003. New York, NY.
[66]Kimmel R.,Fast edge integration, In Level Set Methods and Their Applications in Computer Vision, Chapter 3. Springer Verlag:NY.2003.
[67]A. Tsai, A. Yezzi, and A. S. Willsky. Curve evolution implementation of the Mumford-Shah functional for image segmentation, denoising, interpolation, and magnification, IEEE Tran. Image Process.,10(8)(2001),1169-1186.
[68]Tony F. Chan, B. Yezrielev Sandberg, and Luminita A. Vese, Active Contours without Edges for Vector-Valued Images, Journal of Visual Communication and Image Representation,2(11)(2000),130-141.
[69]Weickert J., Romeny B. and Viergever M., Efficient and Reliable Scheme for Nonlinear Diffusion Filtering,IEEE Transactions on Image Processing,7 (3) (1998), 398-410.
[70]J. Weickert, Applications of nonlinear diffusion in image processing and computer vision. Acta Math. Univ. Comenian. (N.S.),70(1)(2000),33-50.
[71]Bresson X, Esedoglu S, Vandergheynst P, et al. Fast Global Minimization of the Active Contour/Snake Model. J Math Imaging Vis,28(2007),151-167.
[72]Guillermo Sapiro, Vector-Valued Active Contours, IEEE Computer Society Confer-ence on Computer Vision and Pattern Recognition (CVPR'96), (1996):680-685.
[73]Aujol J.-F., Gilboa G., Chan T., Osher S.:Structure-texture image decomposition-modeling, algorithms, and parameter selection. Int. J. Comput. Vis.67(1),111-136 (2006)Moggi E. Notions of computation and monads. Inform Comput,1991,93:55-92.
[74]Chan T., Esedoglu S., Nikolova M., Algorithms for finding global minimizers of image segmentation and denoising models, UCLA CAM Report 04-54 (2004).
© 2004-2018 中国地质图书馆版权所有 京ICP备05064691号 京公网安备11010802017129号 地址：北京市海淀区学院路29号 邮编：100083 电话：办公室：(+86 10)66554848；文献借阅、咨询服务、科技查新：66554700 |