• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Removingmixed noiseinlow ranktexturesbyconvex optimization

    2016-12-14 05:28:37XiaoLiang
    Computational Visual Media 2016年3期

    Xiao Liang()

    ?The Author(s)2016.This article is published with open access at Springerlink.com

    Removingmixed noiseinlow ranktexturesbyconvex optimization

    Xiao Liang1()

    ?The Author(s)2016.This article is published with open access at Springerlink.com

    DOI 10.1007/s41095-016-0056-2 Vol.2,No.3,September 2016,267–276

    This paper introduces a new low rank texture image denoising algorithm,which can restore low rank texture contaminated by both Gaussian and salt-and-pepper noise.The algorithm formulates texture image denoising in terms of solving a low rank matrix optimization problem.Simply assuming low rank is insufficient to describe the properties of natural images,causing high noise amplitudes which lead to unsatisfactory denoising results or serious loss of image details.Thus,in addition to the low rank assumption, the continuity of natural images is also assumed by the algorithm,by adding a total variation regularizer to the optimization objective function. We further give an effective algorithm to solve this optimization problem.By combining the low rank and continuity assumptions,the proposed algorithm overcomes the deficiencies of using either the low rank assumption or total variation regularization alone.Experiments show that our algorithm can effectively remove mixed noise in low rank texture images,and is better than existing algorithms in both its subjective visual effects and in terms of quantitative objective measures.

    image denoising;low rank texture; total variation;convex optimization; augmented Lagrangian method

    1 Introduction

    Image denoising is an extensively studied problem in the image processing community and continues to attract researchers who aim to perform better restoration in the presence of noise. During the past few decades,many intelligent methods havebeen proposed to improve single-image denoising performance. From pixel level filtering methods, such as Gaussian filtering[1],bilateral filtering,and total variation regularization[2],to patch based filtering methods,such as non-local means[3], block-matching 3D filtering(BM3D)[4],and sparse representation [5],single-imagebased denoising performance has been greatly improved,with image details well recovered when the image is slightly noisy.As such filtering methods are widely used in computer vision,work has considered how to speed them up,e.g.,for the bilateral filter[6]and weighted median[7].Comprehensive overviews of image denoising methods can be found in Refs.[8,9].

    Most of the approaches mentioned above consider the input image as an ordinary signal,taking the image as a vector or a set of patches. They do explicitly utilize internal structural information in the image. However,a typical 3D scene of an artificial environment is rich in regular structures. For instance,in an urban environment,the scene istypically filled with man-madeobjectsthat have parallel edges,right-angled corners,regular shapes,symmetric structures,and repeated patterns. Developing algorithms targeted to such images is necessary.In this paper,we mainly study images or textures with low rank structure(see Fig.1 for examples).A rigorous definition[10]of low rank texture is given in Section 2.

    Fig.1 Representative examples of low rank textures.These provide initial images for Figs.3–5.

    Recently,low rank matrix denoising algorithms have been widely studied[11–15].The traditional, image denoising algorithms based on low rank matrix recovery only have the low rank constraint. When Gaussian noise becomes too large,these algorithms produce unsatisfactory denoising results, with serious loss of image details. The reason is partially because that most algorithms use the nuclear norm of the matrix to approximate its rank, in order to get a soluble convex object function, which brings the following problem:as the amplitude of the Gaussian noise increases,the energy of the matrix is compressed more and more seriously.In order to solve the problem of texture image denoising in a mixed noise model,we propose a new algorithm call LRTD(low rank texture denoising)in this paper.

    LRTD formulates texture image denoising in terms of solving a low rank matrix optimization problem.Because the low rank assumption is not sufficient by itself to describe the properties of natural images[16],high noise amplitudes will lead to unsatisfactory denoising results with serious loss of image details.Thus,in addition to the low rank assumption,we also assume image continuity in the algorithm,by adding a total variation regularizer to the optimization objective function.An effective algorithm to solve this optimization problem is also given in the paper. By combining the low rank assumption and the continuity assumption, the proposed LRTD algorithm can overcome the deficiencies of assuming low rank assumption or using total variation regularization alone. The algorithm can effectively remove mixed noise in low rank texture images,and is better than existing algorithms in terms of both subjective visual effects and objective quantitative measurements.

    2 Definition of low rank textures

    In this paper,we consider a 2D texture as a function I0(x,y),defined on ?2.We say that I0is a low rank texture if the family of one-dimensional functions {I0(x,y0)|y0∈?}span a finite low-dimensional linear subspace:

    for some small positive integer k.If r is finite,then we refer to I0as a rank-r texture.Figure 1 shows some ideal low rank textures.To a large extent,the notion of low rank texture unifies many conventional local features.Using this definition,it is easy to see that images of regular symmetric patterns always lead to low rank textures.Thus,the notion of low rank texture encompasses a much broader range of“features”or regions than corners and edges.

    3 Problem formulation

    Before we statistically analyze image denoising,we first define our image formation model:

    where‖E‖0denotes the number of non-zero entries in E,‖·‖F(xiàn)denotes the Frobenius norm,δ>0 is a Gaussian noise intensity parameter,and λ is a weighting parameter which trades off the rank and sparsity of the recovered image.In the above problem,both the rank function and the l0norm can be replaced by convex surrogates[17]:the matrix nuclear norm1The nuclear norm of a matrix is the sum of all its singular values.‖L‖?for rank(L)and the l1norm2The l1norm is the sum of absolute values.‖E‖1for‖E‖0,respectively.Thus,we end up with the following optimization problem:

    Formulation(2)utilizes the low rank nature of the image and the sparsity of the impulsive noise E.But as noted in Ref.[16],while being low rank is a necessary condition for most regular, structured images,it is certainly not sufficient.We need other priors to model additional structures in

    the natural image.Moreover,because the nuclear norm is an approximation to the rank of a matrix, when the noise amplitude is large,formulation(2) leads to over-compression of the nuclear norm[18], causing the total energy of the denoised image to significantly decrease:the picture becomes darker; see for example the fourth column of Fig.5.Thus,to take into account the piecewise smooth continuity of a natural image,we add a total variation regularizer to the optimization problem:

    where‖L‖TV= ‖DxL‖1+‖DyL‖1is the total variation regularizer,in which Dxand Dyare first order forward finite-difference operators in horizontal and vertical directions respectively.Their definitions are

    with periodic boundary conditions;vec(·)represents the vectorization operator.

    4 Denoising by convex optimization

    To solve the convex optimization problem in Eq.(3), we use the alternating direction method(ADM)[19], as it has been proven to be one of the fastest algorithms for solving various low rank matrix completion and recovery problems.To be able to adopt the ADM method to our problem,we need to make our objective function separable. Thus we introduce three auxiliary variables Cx,Cy,and

    W,which turns the optimization problem into the following:

    In formulation(4),the augmented Lagrangian function is defined as

    where L,W,E,Z,Cx,Cy are the unknown variables,Y1,Y2,Y3,Y4are Lagrange multipliers, andμ>0 is a penalty parameter;〈·,·〉indicates inner product.The resulting classic ADM iteration scheme for our problem is given by

    where ρ>1 is a constant. We now focus on efficiently solving the first six steps of the above iterative scheme.

    1)Solving Eq.(6)

    in which UΣVTistheSVD (singularvalue decomposition)of X,and T[·]represents the softthresholding operator defined for scalars as follows:

    for ε≥0;it is extended to vectors and matrices by applying it elementwise.

    2)Solving Eqs.(7)–(9)

    Each of these three variables has closed form solutions,as follows:

    3)Solving Eq.(10)

    Here W also has a closed form solution:

    where Id is the identity matrix.Then we use Fourier transform to solve W[2]:

    where F denotes the 2D Fourier transform operator. The denominator on the right hand side of Eq.(12)is independent of the iteration number k,and so can be precalculated outside the main loop.Therefore,the complexity of solving Eq.(12)is the complexity of one 2D Fourier transform and one inverse 2D Fourier transform.

    4)Solving Eq.(11)

    Following Ref.[13],we write

    Algorithm 1 gives pseudocode of the overall LRTD algorithm.

    5 Results

    In this section we compare our LRTD algorithm with existing approaches.All experiments were performed using MATLAB on a laptop with a 2.30GHz processor and 8GB of RAM.

    We select a set of parameters with the best overall performance for λ and α in our algorithm;LRTD is not sensitive to parameterμ0.ρ=max(1.4?

    σ/600,1.2)is related to noise intensity σ. The greater the σ,the smaller the ρ.

    Algorithm 1:ADM algorithm for solving problem(4) Input:Input image I∈?m×n,parameters λ>0,α>0. Initialize:k=0,L0=I,E0=0,Z0=0,W0=0,Cx=DxI,Cy=DyI,Y1,0=0,Y2,0=0,Y3,0=0,Y4,0=0, μ0>0,ρ>1. WHILE‖Lk+1?Lk‖2/‖Lk‖2≥tolerance DO Lk+1=S(μk)?1?I?Ek?Zk?1μkY3,k+Wk?1μkY4,k/2)?; Cxk+1=Tα/μk(DxWk?Y1,k/μk); Cyk+1=Tα/μk(DyWk?Y1,k/μk); Ek+1=Tλ/μk(I?Lk+1?Zk?Y3,k/μk); Wk+1=F?1?F[DTx(Cxk+1+Y1,k/μk)+DTy(Cyk+1+Y2,k/μk)+L?Y4,k/μk] F[Id+DTxDx+DTyDy] ?;‖N‖F(xiàn) N; Y1,k+1=Y1,k+μk·(Cxk+1?DxLk+1); Y2,k+1=Y2,k+μk·(Cyk+1?DyLk+1); Y3,k+1=Y3,k+μk·(Lk+1+Ek+1+Zk+1?I); Y4,k+1=Y4,k+μk·(Lk+1?Wk+1); μk+1=ρμk; END WHILE Output:Solution(L,E,W,Z,Cx,Cy)to problem(4). Zk+1=min{‖N‖F(xiàn),δ}

    5.1 Comparison with othermixed noise removal methods

    In this paper we consider image denoising problems with mixed noise,in which the image is contaminated by both Gaussian white noise and salt-and-pepper noise. Somewellknown denoising methods such as BM3D work very well to restore images contaminated by pure Gaussian noise,but are completely unable to deal with salt-and-pepper noise[21].Thus,we only compare our LRTD method with three other noise removal methods[13,22,23] specifically designed for mixed noise.

    Following their papers,the parameter settings used were:for Ref.[22],β2=0.00002,tol=10?4, η=1;for Ref.[23],outPer=sr,blocksize=[8,8], stepsize=[2,2];for Ref.[13],η=1.3,β=0.13β0. Besides visual comparison of the results,peak signal to noise ratio(PSNR)was measured to quantitatively evaluate the quality of the restoration results.Given an image L?∈[0,255]m×n,the PSNR of its estimated L is defined as

    A quantitative comparison is shown in Fig.2. Here,we used the image in Fig.1(c)for testing. The percentage of salt-and-pepper noise pixels in the image is denoted by sr(salt-and-pepper noise ratio), while the standard deviation of the white Gaussian noise is denoted by σ.Figure 3(a)shows how PSNR varies for the denoised images as the salt-and-pepper noise ratio sr varies from 0 to 100%with fixed Gaussian noise σ=10;Fig.3(b)shows how PSNR of denoised images varies as the standard deviation of the Guassian noise σ varies from 0 to 60 with a fixed salt-and-pepper noise percentage sr=10%.

    Further quantitative results for the four algorithms were obtained using input images generated by adding salt-and-pepper noise with different levels (sr=15%,30%,45%)mixed with Gaussian noise with different levels(σ=5,15,30,60)to the textures shown in Fig.1.The PSNR values of the restoration results of these methods are summarized in Table 1. Figures 3–5 give qualitative comparisons;the original images in these experiment are shown in Fig.1.We can see from these results that LRTD works better on low rank texture images than previous algorithms.

    From Fig.2 and Table 1 we can see that LRTD's ability to process salt-and-pepper noise is very good. In low Gaussian noise environments(σ<10),as the percentage of salt-and-pepper noise increases from 0%to 60%,the PSNRs of our image restoration results do not decrease significantly,and are always more than 30dB.However,the LRTD algorithm is more sensitive to Gaussian noise.As Gaussian noise increases to about σ=60,results shown in Fig.5 display the excessive compression issue caused by use of convex surrogates for the rank function.Although the structure of the restored image is still quite good, due to the compression of the overall energy of the input image during the optimization process,the resulting PSNR decreases significantly.

    Fig.2 Variation of PSNR with varying amounts of salt-and-pepper noise and Gaussian noise.(a)σ=10,sr varying from 0 to 100%;(b) sr=10%,σ varying from 0 to 60.

    Table 1 Comparison with other mixed noise removal methods,showing PSNR values for varying amounts of salt-and-pepper and Gaussian noise

    The results in this subsection demonstrate that our LRTD denoising method can effectively remove mixed noise in low rank texture images,and works better than other existing algorithms.Addition of the TV regularizer to the optimization objective function has a good effect on avoiding the problem of excessive compression.Our LRTD shows significant improvement compared to the simple low rank optimization algorithm[13].

    5.2 Computation time

    We performed a further experiment to test the convergence performance of LRTD.Let I=L?+E?+Z?be the noisy data matrix,where L?and E?are the low rank and sparse components to be recovered. We generated a rank 2 checkerboard image as L?∈?320×320.The support ? of the impulsive noise E?(sparse but large)was chosen uniformly at random, and the non-zero entries of E?were i.i.d.uniform in the interval[?500,500].

    As our aim was to test running speed and convergence of LRTD,we set σ=0;this differs from the mixed noise used in the earlier tests as there is no Gaussian noise.In this case,the algorithm can converge to the ground truth noiseless matrix L?.

    Here we compare our method with another low rank matrix recovery algorithm ASALM[13].The objective function of ASALM is given in Eq.(2):the only difference between ASALM with our LRTD is that ASALM lacks the TV regularizer.Following their paper,we set the parameters for ASALM to η=1.3,β=0.13β0.Table 2 shows that LRTD is slower than ASALM.However,LRTD performs better in denoising.We believe the trade-off to be acceptable.

    Table 2 Computation time and number of iterations

    Fig.3 Qualitative comparison of results of various methods(1).Left to right:input noisy image with σ=10,sr=10%,denoising results of Ref.[22],KALS[23],ASALM[13],and our denoising result.

    6 Conclusions and discussion

    This paper introduced a new low rank texture image denoising algorithm,which can restore low rank texture contaminated by both Gaussian and salt-and-pepper noise. Our method directly uses raw pixel values of the image as the matrix and models texture image denoising as a low rank matrix optimization problem. Besides the low rank assumption,we also utilize the assumption of continuity of natural images,by adding a total variation regularizer to the optimization objective function. Our results demonstrate that the TV regularizer indeed helps low rank texture denoising.

    Through extensive experiments,we have shown that our LRTD method works better than existing algorithms both in subjective visual terms and through objective quantitative measures.

    Although our algorithm works robustly under a broad range of conditions and for a wide range of regular textures,it may fail if the conditions are too challenging or the assumptions are violated.Figure 6 shows some such cases.The denoising results in Fig.6(c)and Fig.6(e)are reasonable but not perfect. As mentioned earlier in Section 1,our algorithm is not designed to work on random textures.Although there has been work in the literature showing that it is possible to get a reasonable denoised result for

    such examples as grass lawns,our algorithm is not designed to handle such cases.LRTD is effective for regular symmetric textures,but not for random textures which normally have high rank matrices.

    Fig.4 Qualitative comparison of results of various methods(2).Left to right:input noisy image with σ=15,sr=45%,denoising results of Ref.[22],KALS[23],ASALM[13],and our denoising result.

    Clearly,a natural image of low rank texture may be deformed by the camera projection and undergoes a certain domain transformation (say affine or projective). The transformed texture,viewed as a matrix,in general is no longer low rank in the image domain.Nevertheless,by utilizing advanced convex optimization tools[10]from matrix rank minimization,we can recover a low rank texture from the deformed image and the associated deformation. As TILT(transform invariant low rank texture)[10] uses image interpolation during its iterations,simply running LRTD on TILT results would hurt image performance.Finding how to combine TILT with LRTD to enable a wider range of applications would be valuable future work.

    Acknowledgements

    We sincerely appreciate Zhouchen Lin and Xin Tong's help with valuable suggestions and comments for this paper.

    [1]Rank,K.;Unbehauen,R.An adaptive recursive 2-D filter for removal of Gaussian noise in images.IEEE

    Transactions on Image Processing Vol.1,No.3,431–436,1992.

    Fig.5 Qualitative comparison of results of various methods(3).Left to right:input noisy image with σ=60,sr=10%,denoising results of Ref.[22],KALS[23],ASALM[13],and our denoising result.

    Fig.6 Examples with decreasing regularity and increasing randomness.(a),(c),and(e)input noisy images,with σ=10,sr=10%;(b), (d),and(f)denoising results and their PSNRs.

    [2]Chan,S.H.;Khoshabeh,R.;Gibson,K.B.;Gill,P.E.; Nguyen,T.Q.An augmented Lagrangian method for total variation video restoration.IEEE Transactions on Image Processing Vol.20,No.11,3097–3111,2011.

    [3]Buades,A.;Coll,B.;Morel,J.-M.A non-local algorithm for image denoising.In:Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition,Vol.2,60–65,2005.

    [4]Dabov,K.;Foi,A.;Katkovnik,V.;Egiazarian, K.Image denoising by sparse 3-D transform-domain

    collaborative filtering.IEEE Transactions on Image Processing Vol.16,No.8,2080–2095,2007.

    [5]Afonso,M.V.;Sanches,J.M.R.Blind inpainting using l0and total variation regularization.IEEE Transactions on Image Processing Vol.24,No.7, 2239–2253,2015.

    [6]Gastal, E.S.L.; Oliveira, M.M.Adaptive manifoldsforreal-timehigh-dimensionalfiltering. ACM Transactions on Graphics Vol.31,No.4,Article No.33,2012.

    [7]Zhang,Q.;Xu,L.;Jia,J.100+ timesfaster weighted median filter(WMF).In:Proceedings of IEEE Conference on Computer Vision and Pattern Recognition,2830–2837,2014.

    [8]Buades,A.;Coll,B.;Morel,J.-M.A review of image denoising algorithms,with a new one.SIAM Journal on Multiscale Modeling and Simulation: A SIAM Interdisciplinary Journal Vol.4,No.2,490–530,2005.

    [9]Chatterjee,P.;Milanfar,P.Is denoising dead?IEEE Transactions on Image Processing Vol.19,No.4,895–911,2010.

    [10]Zhang,Z.;Ganesh,A.;Liang,X.;Ma,Y.TILT: Transform-invariant low-rank textures.International Journal of Computer Vision Vol.99,No.1,1–24,2012.

    [11]Dong,W.;Shi,G.;Li,X.Nonlocal image restoration with bilateralvariance estimation: A low-rank approach.IEEE Transactions on Image Processing Vol.22,No.2,700–711,2013.

    [12]Shabalin,A.A.;Nobel,A.B.Reconstruction of a lowrank matrix in the presence of Gaussian noise.Journal of Multivariate Analysis Vol.118,No.5,67–76,2013.

    [13]Tao,M.;Yuan,X.Recovering low-rank and sparse components of matrices from incomplete and noisy observations.SIAM Journal on Optimization Vol.21, No.1,57–81,2011.

    [14]Zhang,Y.;Liu,Y.;Li,X.;Zhang,C.Salt and pepper noise removal in surveillance video based on low-rank matrix recovery.Computational Visual Media Vol.1, No.1,59–68,2015.

    [15]Zhou,Z.;Li,X.;Wright,J.;Cand`es,E.J.;Ma,Y. Stable principal component pursuit.In:Proceedings of IEEE International Symposium on Information Theory,1518–1522,2010.

    [16]Liang,X.;Ren,X.;Zhang,Z.;Ma,Y.Texture repairing by unified low rank optimization.Journal of Computer Science and Technology Vol.31,No.3,525–546,2016.

    [17]Cand`es,E.J.;Li,X.;Ma,Y.;Wright,J.Robust principal component analysis?Journal of the ACM Vol.58,No.3,Article No.11,2011.

    [18]Wang,Z.;Zhang,J.;Chen,G.Mixture noise image denoising using reweighted low-rank matrix recovery. Compuer Science Vol.43,No.1,298–301,2016.(in Chinese)

    [19]Lin,Z.;Ganesh,A.;Wright,J.;Wu,L.;Chen,M.; Ma,Y.Fast convex optimization algorithms for exact recovery of a corrupted low-rank matrix.Technical Report.University of Illinois at Urbana-Champaign, 2009.

    [20]Cand`es,E.J.;Plan,Y.Matrix completion with noise. Proceedings of the IEEE Vol.98,No.6,925–936,2010.

    [21]Djurovi′c,I.BM3D filter in salt-and-pepper noise removal.EURASIP Journal on Image and Video Processing Vol.2016,13,2016.

    [22]Cai,J.-F.;Chan,R.H.;Nikolova,M.Fast twophase image deblurring under impulse noise.Journal of Mathematical Imaging and Vision Vol.36,46–53, 2010.

    [23]Wang,Y.;Szlam,A.;Lerman,G.Robust locally linear analysis with applications to image denoising and blind inpainting.SIAM Journal on Imaging Sciences Vol.6, No.1,526–562,2013.

    Xiao Liang iscurrently a Ph.D. student in computer science and technology at the Institute for Advanced Study in Tsinghua University,Beijing, China. Her adviser is Prof. Harry Shum.She received her B.E.degree in electronic engineering from Tsinghua University. During herstudy,she interned at Microsoft Research Asia for over four years. Her research interests include texture processing,3D computervision and sparsity,and low rank matrix recovery.

    Open Access The articles published in this journal are distributed under the terms of the Creative Commons Attribution 4.0 International License(http:// creativecommons.org/licenses/by/4.0/), which permits unrestricted use,distribution,and reproduction in any medium,provided you give appropriate credit to the original author(s)and the source,provide a link to the Creative Commons license,and indicate if changes were made.

    Other papers from this open access journal are available free of charge from http://www.springer.com/journal/41095. To submit a manuscript,please go to https://www. editorialmanager.com/cvmj.

    1 Institute for Advanced Study,Tsinghua University, Beijing100084,China. E-mail: liangx04@mails. tsinghua.edu.cn().

    Manuscript received:2016-04-22;accepted:2016-05-25

    99九九在线精品视频| 天天影视国产精品| 欧美人与善性xxx| 大话2 男鬼变身卡| 黑人高潮一二区| 亚洲欧美日韩卡通动漫| 熟妇人妻不卡中文字幕| 久久久久久久久久久免费av| 26uuu在线亚洲综合色| 天堂8中文在线网| 久久久久国产精品人妻一区二区| 日韩视频在线欧美| 中文字幕人妻丝袜制服| av又黄又爽大尺度在线免费看| 精品国产国语对白av| 精品99又大又爽又粗少妇毛片| 人人妻人人爽人人添夜夜欢视频| 国产精品无大码| 亚洲欧洲日产国产| 久久久久国产精品人妻一区二区| 午夜福利,免费看| 亚洲精品乱码久久久久久按摩| 久久人人爽人人爽人人片va| 下体分泌物呈黄色| 18+在线观看网站| 两个人的视频大全免费| 久久久精品94久久精品| 麻豆精品久久久久久蜜桃| 99久久精品一区二区三区| 亚洲av日韩在线播放| 桃花免费在线播放| 国内精品宾馆在线| 亚洲精品久久成人aⅴ小说 | 精品卡一卡二卡四卡免费| 欧美日韩亚洲高清精品| 欧美激情极品国产一区二区三区 | 久久久久久久精品精品| 国产一级毛片在线| 亚洲av二区三区四区| 精品久久久久久久久亚洲| 黄色怎么调成土黄色| 亚洲欧美日韩另类电影网站| 国产成人一区二区在线| av播播在线观看一区| 国产片内射在线| 女性被躁到高潮视频| 亚洲欧美色中文字幕在线| 大陆偷拍与自拍| 九九久久精品国产亚洲av麻豆| 91精品伊人久久大香线蕉| 18禁观看日本| 国产精品一区二区在线观看99| 又大又黄又爽视频免费| 99久久综合免费| 亚洲不卡免费看| 久久久国产精品麻豆| 日本wwww免费看| 王馨瑶露胸无遮挡在线观看| 免费少妇av软件| 精品熟女少妇av免费看| 久久久国产一区二区| 亚洲国产色片| a级片在线免费高清观看视频| 国产黄色免费在线视频| 一区二区三区四区激情视频| 午夜精品国产一区二区电影| 插逼视频在线观看| 一级a做视频免费观看| 丝袜脚勾引网站| 精品一品国产午夜福利视频| 免费人妻精品一区二区三区视频| 日日摸夜夜添夜夜爱| 久久久久久久大尺度免费视频| 亚洲在久久综合| 日产精品乱码卡一卡2卡三| 日韩强制内射视频| 久热久热在线精品观看| 黄色一级大片看看| 亚洲国产精品专区欧美| 一级毛片我不卡| 一个人免费看片子| a级毛片在线看网站| 亚洲精品国产色婷婷电影| 亚洲欧美一区二区三区国产| 18+在线观看网站| 日产精品乱码卡一卡2卡三| 人人澡人人妻人| 国产69精品久久久久777片| 国产不卡av网站在线观看| 超色免费av| 人人澡人人妻人| 如何舔出高潮| 美女脱内裤让男人舔精品视频| 亚洲av.av天堂| 亚洲精品国产av成人精品| 丝袜喷水一区| 美女脱内裤让男人舔精品视频| 97超视频在线观看视频| 国产精品久久久久久精品电影小说| 99久久精品一区二区三区| 亚洲一区二区三区欧美精品| 亚洲,一卡二卡三卡| 国产深夜福利视频在线观看| 国产免费福利视频在线观看| 91久久精品国产一区二区三区| 国产精品国产三级国产专区5o| 蜜桃在线观看..| 免费高清在线观看视频在线观看| 国精品久久久久久国模美| 久久久久久久久久久丰满| 色网站视频免费| 黑人高潮一二区| av女优亚洲男人天堂| 80岁老熟妇乱子伦牲交| 国产精品成人在线| 看十八女毛片水多多多| 亚洲一级一片aⅴ在线观看| 日韩亚洲欧美综合| 免费人成在线观看视频色| 交换朋友夫妻互换小说| 综合色丁香网| 国产爽快片一区二区三区| 一级毛片黄色毛片免费观看视频| 国产一区亚洲一区在线观看| 成年人午夜在线观看视频| 久热这里只有精品99| 久久久午夜欧美精品| 国产成人av激情在线播放 | 亚洲av国产av综合av卡| 在线观看美女被高潮喷水网站| 精品99又大又爽又粗少妇毛片| 国产伦理片在线播放av一区| 日韩制服骚丝袜av| 国产精品一区www在线观看| 欧美老熟妇乱子伦牲交| 国产成人91sexporn| 99久国产av精品国产电影| 久热这里只有精品99| 在线观看免费日韩欧美大片 | 肉色欧美久久久久久久蜜桃| 成人影院久久| 99久久人妻综合| 亚洲第一区二区三区不卡| 免费av中文字幕在线| 国产精品一二三区在线看| 欧美激情国产日韩精品一区| 久久久久视频综合| 国产视频内射| 国产高清不卡午夜福利| 男女国产视频网站| 欧美成人午夜免费资源| 久久久亚洲精品成人影院| 丰满迷人的少妇在线观看| 考比视频在线观看| 亚洲激情五月婷婷啪啪| 卡戴珊不雅视频在线播放| 丝袜喷水一区| 哪个播放器可以免费观看大片| 丰满少妇做爰视频| 久久av网站| 国产伦精品一区二区三区视频9| 一本大道久久a久久精品| 欧美变态另类bdsm刘玥| 国产伦精品一区二区三区视频9| 中文乱码字字幕精品一区二区三区| 国产伦理片在线播放av一区| 国产一区亚洲一区在线观看| 999精品在线视频| 久久国产精品大桥未久av| 少妇的逼好多水| 国产亚洲一区二区精品| 性高湖久久久久久久久免费观看| 国产毛片在线视频| 老司机影院毛片| 亚洲av成人精品一二三区| 2021少妇久久久久久久久久久| 日韩欧美一区视频在线观看| 草草在线视频免费看| 免费日韩欧美在线观看| 国产欧美亚洲国产| 欧美少妇被猛烈插入视频| 妹子高潮喷水视频| 国产成人精品无人区| 免费大片18禁| 亚洲国产毛片av蜜桃av| 欧美日韩视频精品一区| 亚洲精品国产av成人精品| 午夜免费观看性视频| 黄片无遮挡物在线观看| a 毛片基地| 国产精品99久久久久久久久| 国产国语露脸激情在线看| 在线 av 中文字幕| 大香蕉久久网| 男女高潮啪啪啪动态图| 久久久久久久久久久丰满| 成人毛片a级毛片在线播放| 欧美性感艳星| 日韩电影二区| 国产一区二区三区av在线| 99九九在线精品视频| 赤兔流量卡办理| 欧美bdsm另类| 99热6这里只有精品| 亚洲人成77777在线视频| 亚洲欧洲日产国产| 国产亚洲精品第一综合不卡 | 多毛熟女@视频| 如何舔出高潮| 亚洲av中文av极速乱| 亚洲婷婷狠狠爱综合网| 尾随美女入室| 日韩中文字幕视频在线看片| 亚洲av不卡在线观看| 狠狠婷婷综合久久久久久88av| 欧美另类一区| 伊人久久国产一区二区| 亚洲精品自拍成人| 国产成人免费无遮挡视频| 99久久精品一区二区三区| 3wmmmm亚洲av在线观看| 国产精品一二三区在线看| 99视频精品全部免费 在线| 免费av中文字幕在线| 三上悠亚av全集在线观看| 成年美女黄网站色视频大全免费 | 亚洲性久久影院| 色吧在线观看| 日本黄色片子视频| 国产女主播在线喷水免费视频网站| 美女xxoo啪啪120秒动态图| 亚洲不卡免费看| 91午夜精品亚洲一区二区三区| 毛片一级片免费看久久久久| 久久久久人妻精品一区果冻| 久久狼人影院| 十分钟在线观看高清视频www| a级毛色黄片| 一级毛片黄色毛片免费观看视频| 午夜福利视频精品| 伦精品一区二区三区| 亚洲av福利一区| 亚洲国产欧美日韩在线播放| 18在线观看网站| 国产在视频线精品| 国产精品一区二区在线观看99| 亚洲精品国产色婷婷电影| 亚洲精品一区蜜桃| 日韩中字成人| 精品人妻熟女av久视频| 欧美xxⅹ黑人| 一区二区三区免费毛片| 91精品三级在线观看| 日本猛色少妇xxxxx猛交久久| 中文字幕亚洲精品专区| 亚洲精品自拍成人| 欧美日韩av久久| 国产白丝娇喘喷水9色精品| a级毛片免费高清观看在线播放| 高清在线视频一区二区三区| 国产一区二区在线观看日韩| 日本黄色片子视频| freevideosex欧美| 亚洲国产精品成人久久小说| 欧美日韩综合久久久久久| 亚洲情色 制服丝袜| 国产色婷婷99| 精品一区二区免费观看| 在线亚洲精品国产二区图片欧美 | 香蕉精品网在线| 天天操日日干夜夜撸| 99久久综合免费| 日韩成人伦理影院| 男女国产视频网站| 免费人成在线观看视频色| 色婷婷av一区二区三区视频| 青春草国产在线视频| 中文字幕人妻丝袜制服| 国产免费一区二区三区四区乱码| 久久久久久久久久久免费av| 国产黄片视频在线免费观看| 高清视频免费观看一区二区| 久久ye,这里只有精品| 久久久久久久亚洲中文字幕| 亚洲欧美成人综合另类久久久| 久久精品国产亚洲网站| 22中文网久久字幕| 午夜福利影视在线免费观看| 99国产综合亚洲精品| 久久久久久久精品精品| 99re6热这里在线精品视频| 免费久久久久久久精品成人欧美视频 | 欧美日韩国产mv在线观看视频| 午夜福利,免费看| 九色成人免费人妻av| 91成人精品电影| 久久精品国产自在天天线| 建设人人有责人人尽责人人享有的| 大话2 男鬼变身卡| 高清欧美精品videossex| 女性生殖器流出的白浆| 亚洲精品av麻豆狂野| 久久亚洲国产成人精品v| xxx大片免费视频| 久久精品国产鲁丝片午夜精品| 亚洲美女黄色视频免费看| 久久久久精品性色| 欧美xxⅹ黑人| 一本大道久久a久久精品| 波野结衣二区三区在线| 亚洲欧美日韩卡通动漫| 日本黄色日本黄色录像| 国产高清有码在线观看视频| 尾随美女入室| 国产精品女同一区二区软件| 久久久欧美国产精品| 高清av免费在线| 国产亚洲av片在线观看秒播厂| 日韩伦理黄色片| 国产精品久久久久久精品古装| 一级,二级,三级黄色视频| 成年人午夜在线观看视频| 99精国产麻豆久久婷婷| 嫩草影院入口| 欧美激情极品国产一区二区三区 | 最近的中文字幕免费完整| 日本免费在线观看一区| 欧美日韩国产mv在线观看视频| 伦理电影大哥的女人| 色视频在线一区二区三区| 久久亚洲国产成人精品v| 欧美少妇被猛烈插入视频| 国产免费又黄又爽又色| 国产成人精品婷婷| 男人添女人高潮全过程视频| 一级片'在线观看视频| 国产一级毛片在线| 3wmmmm亚洲av在线观看| 女的被弄到高潮叫床怎么办| 久久免费观看电影| 亚洲人与动物交配视频| 男女边摸边吃奶| 曰老女人黄片| 中文天堂在线官网| 久久精品夜色国产| 亚洲精品久久久久久婷婷小说| 在线 av 中文字幕| 母亲3免费完整高清在线观看 | 国产精品免费大片| 亚洲精品久久成人aⅴ小说 | 在线观看一区二区三区激情| 日韩大片免费观看网站| 亚洲国产毛片av蜜桃av| 久久人人爽人人爽人人片va| 男人添女人高潮全过程视频| 午夜久久久在线观看| 夜夜爽夜夜爽视频| 夫妻午夜视频| 久久久欧美国产精品| 高清av免费在线| 久久国产精品男人的天堂亚洲 | 观看美女的网站| 中国美白少妇内射xxxbb| 久久影院123| 在线观看免费视频网站a站| 欧美精品国产亚洲| 高清在线视频一区二区三区| 爱豆传媒免费全集在线观看| 男女无遮挡免费网站观看| 精品久久蜜臀av无| 精品一品国产午夜福利视频| 黑人猛操日本美女一级片| 考比视频在线观看| 久久影院123| 国产伦理片在线播放av一区| 最近中文字幕高清免费大全6| 亚洲,欧美,日韩| 天美传媒精品一区二区| 成年人免费黄色播放视频| 欧美性感艳星| 九草在线视频观看| 国产亚洲最大av| 国产又色又爽无遮挡免| 美女xxoo啪啪120秒动态图| 国产成人一区二区在线| 最近中文字幕高清免费大全6| 亚洲国产日韩一区二区| 亚洲,一卡二卡三卡| 精品少妇黑人巨大在线播放| 日本爱情动作片www.在线观看| 看免费成人av毛片| 高清不卡的av网站| 色网站视频免费| 日韩欧美精品免费久久| 看十八女毛片水多多多| 能在线免费看毛片的网站| 插逼视频在线观看| 成人无遮挡网站| 中文字幕人妻丝袜制服| 2021少妇久久久久久久久久久| 亚洲第一区二区三区不卡| 热99国产精品久久久久久7| 精品熟女少妇av免费看| 黄色视频在线播放观看不卡| freevideosex欧美| 妹子高潮喷水视频| 观看av在线不卡| 久久久久久人妻| 欧美bdsm另类| 中文字幕人妻熟人妻熟丝袜美| 亚州av有码| 又黄又爽又刺激的免费视频.| 国产一区亚洲一区在线观看| 午夜免费观看性视频| 狂野欧美激情性xxxx在线观看| 国产在线一区二区三区精| 日韩精品有码人妻一区| 老女人水多毛片| 五月玫瑰六月丁香| av视频免费观看在线观看| a级毛片免费高清观看在线播放| 亚洲精华国产精华液的使用体验| 久久久精品94久久精品| 伦精品一区二区三区| 黄色毛片三级朝国网站| 99视频精品全部免费 在线| 国内精品宾馆在线| 男女边摸边吃奶| 妹子高潮喷水视频| 又大又黄又爽视频免费| 在线观看人妻少妇| 欧美少妇被猛烈插入视频| 97超碰精品成人国产| xxxhd国产人妻xxx| 五月开心婷婷网| 精品久久久精品久久久| 久久99一区二区三区| 亚洲人与动物交配视频| 如日韩欧美国产精品一区二区三区 | 丁香六月天网| 又粗又硬又长又爽又黄的视频| 久久这里有精品视频免费| 成人影院久久| 在线播放无遮挡| 国产综合精华液| 欧美激情极品国产一区二区三区 | 国产在线一区二区三区精| 中国三级夫妇交换| av视频免费观看在线观看| 国产成人av激情在线播放 | 日韩强制内射视频| 99国产精品免费福利视频| 少妇猛男粗大的猛烈进出视频| 国产熟女欧美一区二区| 国产精品99久久99久久久不卡 | 黑人巨大精品欧美一区二区蜜桃 | 女的被弄到高潮叫床怎么办| 人妻夜夜爽99麻豆av| 丰满少妇做爰视频| av有码第一页| 久久人人爽av亚洲精品天堂| 伦精品一区二区三区| 久久国产亚洲av麻豆专区| av.在线天堂| 精品人妻熟女av久视频| av免费观看日本| 欧美3d第一页| 七月丁香在线播放| 欧美变态另类bdsm刘玥| 免费观看性生交大片5| 99视频精品全部免费 在线| 精品一区二区三区视频在线| 久久99精品国语久久久| 亚洲精品乱码久久久v下载方式| 日韩大片免费观看网站| 国产精品人妻久久久久久| 黄色毛片三级朝国网站| 日本-黄色视频高清免费观看| 最近的中文字幕免费完整| 如日韩欧美国产精品一区二区三区 | 国产白丝娇喘喷水9色精品| 夜夜骑夜夜射夜夜干| 久久久亚洲精品成人影院| 飞空精品影院首页| 精品久久蜜臀av无| 尾随美女入室| 一级毛片我不卡| 男女边摸边吃奶| 国产精品蜜桃在线观看| 在线观看三级黄色| 在线播放无遮挡| 嫩草影院入口| 大话2 男鬼变身卡| 国产69精品久久久久777片| 极品人妻少妇av视频| 亚洲精品久久成人aⅴ小说 | 亚洲精品一二三| 亚洲精品av麻豆狂野| 日本黄大片高清| 久久精品久久久久久久性| 国产精品秋霞免费鲁丝片| h视频一区二区三区| 天堂8中文在线网| 人体艺术视频欧美日本| 午夜激情福利司机影院| av福利片在线| 大又大粗又爽又黄少妇毛片口| 色视频在线一区二区三区| 黑丝袜美女国产一区| 97精品久久久久久久久久精品| 日本午夜av视频| tube8黄色片| 一级二级三级毛片免费看| 波野结衣二区三区在线| 在线免费观看不下载黄p国产| 亚洲色图 男人天堂 中文字幕 | 亚洲欧美成人精品一区二区| 91精品伊人久久大香线蕉| 成人毛片a级毛片在线播放| 蜜桃国产av成人99| 亚洲av福利一区| 欧美+日韩+精品| 一本久久精品| 久久99精品国语久久久| 老女人水多毛片| 9色porny在线观看| 久久久久久久久久人人人人人人| 狂野欧美激情性bbbbbb| 最新中文字幕久久久久| 精品视频人人做人人爽| 精品少妇久久久久久888优播| 亚洲欧美一区二区三区黑人 | 九色亚洲精品在线播放| 日韩中字成人| 日韩一区二区视频免费看| 亚洲色图综合在线观看| 亚洲国产精品一区二区三区在线| 亚洲av综合色区一区| 亚洲国产欧美在线一区| 久久久久国产网址| 一本一本综合久久| 日韩中字成人| 亚洲精品久久成人aⅴ小说 | 久久久久久久久久成人| 成人无遮挡网站| 18在线观看网站| 国产亚洲av片在线观看秒播厂| 国产高清国产精品国产三级| 国产熟女欧美一区二区| 成人二区视频| 中文精品一卡2卡3卡4更新| 91精品三级在线观看| 色婷婷久久久亚洲欧美| 婷婷色综合大香蕉| 日韩一区二区视频免费看| 热re99久久国产66热| 天天操日日干夜夜撸| 熟妇人妻不卡中文字幕| 国国产精品蜜臀av免费| 成人亚洲欧美一区二区av| 久久99热6这里只有精品| 日韩大片免费观看网站| 精品人妻熟女毛片av久久网站| 国产片特级美女逼逼视频| 3wmmmm亚洲av在线观看| 赤兔流量卡办理| av电影中文网址| 国产午夜精品久久久久久一区二区三区| 一级爰片在线观看| 国产日韩一区二区三区精品不卡 | 久久 成人 亚洲| 91久久精品电影网| 久久久久久人妻| 国产高清国产精品国产三级| 免费观看在线日韩| 美女视频免费永久观看网站| 免费久久久久久久精品成人欧美视频 | videos熟女内射| 亚洲欧洲精品一区二区精品久久久 | 五月伊人婷婷丁香| 亚洲欧美色中文字幕在线| 综合色丁香网| 纵有疾风起免费观看全集完整版| 丝瓜视频免费看黄片| 亚洲国产精品一区三区| 中文字幕人妻丝袜制服| 午夜精品国产一区二区电影| 免费av中文字幕在线| 国产有黄有色有爽视频| 欧美 日韩 精品 国产| 午夜久久久在线观看| 大陆偷拍与自拍| 欧美丝袜亚洲另类| 国产精品免费大片| 精品国产一区二区三区久久久樱花| 久久精品久久久久久久性| 久久久国产一区二区| 久久精品人人爽人人爽视色| 建设人人有责人人尽责人人享有的| 中文精品一卡2卡3卡4更新| 精品酒店卫生间| 国产精品国产三级国产专区5o| 777米奇影视久久| 女性生殖器流出的白浆| 国产视频内射| 少妇的逼好多水| 久久女婷五月综合色啪小说| 国产69精品久久久久777片| 日本猛色少妇xxxxx猛交久久| 老司机影院毛片| 亚洲国产欧美日韩在线播放| a级毛片免费高清观看在线播放| 插逼视频在线观看| 纯流量卡能插随身wifi吗| 青春草亚洲视频在线观看|