• <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 ?

    Enhancement of Biomass Material Characterization Images Using an Improved U-Net

    2022-08-24 12:56:42ZuozhengLianHongZhaoQianjunZhangHaizhenWangandErdun
    Computers Materials&Continua 2022年7期

    Zuozheng Lian, Hong Zhao, Qianjun Zhang, Haizhen Wangand E.Erdun

    1College of Computer and Control Engineering, Qiqihar University, Qiqihar, 161006, China

    2College of Light Industry and Textile, Qiqihar University, Qiqihar, 161006, China

    3Elastic Block Store Performance Team, Amazon Web Services, Boston, 02210, USA

    Abstract: For scanning electronmicroscopes with high resolution and a strong electric field, biomass materials under observation are prone to radiation damage from the electron beam.This results in blurred or non-viable images, which affect further observation of material microscopic morphology and characterization.Restoring blurred images to their original sharpness is still a challenging problem in image processing.Traditionalmethods can’t effectively separate image context dependency and texture information, affect the effect of image enhancement and deblurring, and are prone to gradient disappearance during model training, resulting in great difficulty in model training.In this paper,we propose the use of an improvedU-Net (U-shapedConvolutional Neural Network) to achieve image enhancement for biomass material characterization and restore blurred images to their original sharpness.The main work is as follows: use of depthwise separable convolution instead of standard convolution in U-Net to reduce model computation effort and parameters; embedding wavelet transform into the U-Net structure to separate image context and texture information, thereby improving image reconstruction quality; using dense multi-receptive field channel modules to extract image detail information, thereby better transmitting the image features and network gradients, and reduce the difficulty of training.The experiments show that the improved U-Net model proposed in this paper is suitable and effective for enhanced deblurring of biomass material characterization images.The PSNR (Peak Signal-to-noise Ratio) and SSIM (Structural Similarity) are enhanced as well.

    Keywords: U-Net; wavelet transform; image enhancement; biomass material characterization

    1 Introduction

    In recent years, a looming energy crisis and greater emphasis on environmental protection have led to 100% environmentally-friendly biomass composites becoming a popular research topic.Their sizes have entered the submicron level, which brings about some difficulties when measuring and characterizing nanoscale surfaces [1].The rapid development of scanning electronmicroscopy imagingis now a common and effective analytical technique used for surface morphology observation.It has high resolution, high magnification, strong depth of field, a large field of view, stereoscopic imaging and can directly observe microstructures on uneven surfaces of various specimens [2].However, when using conventional electron microscopy for direct observation of plant fiber samples, the high-energy electron beam can penetrate the sample, blurring its edges and making it impossible to capture images in severe cases [3].The issue of unclear material characterization images due to irradiation damage or breakdown of the SEM electric field affecting biomass samples is a common phenomenon when researching on the characterization of natural fiber composites.

    All types of biomass composites have their own unique microstructures, and new imaging techniques have played an important role in early studies of hemp-plastic composites.However, there is a tendency for the electron beam to radiate damage to biomass materials under the high-resolution strong electric field during scanning electron microscopy.This leads to blurred or non-viable images,which affect further observation of microscopic morphology.Therefore, there is an urgent need to develop a method for extracting high-resolution and high-magnification microscopic morphological images in spite of sample damage.This damage limits the use of equipment, resulting in low-quality blurred images from electron microscopy which affects the analysis of samples.The restoration of blurred material images to their original sharpness is of great importance.However, due to great variability in the structural complexity and functionality ofmaterials, enhanced sharpening using deep learning techniques for natural fiber composite characterization images can reduce the experimental workload of researchers.This can also facilitate observation of material surface morphology, provide a convenient way to study the morphological structure of samples and help to improve material properties and optimize the preparation process.

    Therefore, in this work, we propose an improved U-Net for sharpening biomass material characterization images.First, depthwise separable convolution is used instead of standard convolution in the encoder of U-Net to reduce the computation workload and the number of parameters.In addition, a dense multi-receiver domain channel module in the deep layer decoder of U-Net is used to extract image detail information to improve image quality.Second, wavelet transform and inverse wavelet transform are embedded into U-Net to reduce model training difficulty and the loss of image information.Third, contextual channel attention modules are used to improve the skip connection and the decode connection for contextual information fusion.The experiments show that the improved U-Net proposed in this paper is suitable and effective for the sharpening of biomass material characterization images.The peak signal-to-noise ratio and structural similarity are enhanced as well.

    The remainder of this paper is organized as follows.In Section 2, we briefly introduce the backgroud of the related work about material characterization and image enhancement.The details of improved U-Net model proposed and the method of image deblurring are described in Section 3.In Section 4, we show the details of the experimental settings and results.Finally, the conclusions and our further works are given in Section 5.

    2 Related Work

    2.1 Material Characterization

    In the field of materials science, SEM (Scanning Electron Microscope) is often used to receive the information beam of the secondary electron beam [4] to form an image of the observed sample.The properties of the observed sample are obtained by analysis of the image.However, due to objective factors such as charge effects, edge effects, sample damage, poor conductivity of the sample and subjective factors such as the skill of the operator, SEM is prone to producing low-quality images.These low-quality images have defects such as blurring, low contrast, high noise and inconspicuous edges, which make it difficult to extract sample information effectively and hinder scientific research.

    Most research has focused on the improvement of SEM operational methods [5] and the accuracy of SEM instrument design and fabrication.The former approach relies on the proficiency of the operator, while the latter approach incurs higher costs.In the Information Age, digital image processing techniques [6] have matured and are widely used in various fields.SEM images can also be enhanced using digital image processing methods.Yu et al.[7] designed a suitable enhancement algorithm for SEM images to improve image quality and further designed a method to outline the edges of the sample so as to extend the SEM image analysis function.Xu et al.[8] improved the quality of SEM images by a spatial domain image enhancement method to improve resolution and adjust for the appropriate brightness.The image recognition function was further investigated to automatically outline sample edges, which makes it easy for researchers to identify sample objects precisely and with ease.

    2.2 Image Enhancement

    Most of the existing traditional image enhancement sharpening methods use regularization and hand-crafted prior images to estimate the blur kernel, followed by iterative optimization to gradually restore a clear image.The traditional approach involves the estimation of complex blur kernels, which leads to a tedious sharpening process, poor real-time performance and poor algorithm performance.With the rapid development of deep learning, image sharpening methods based on deep learning [9]have been increasingly applied.Sun et al.[10] used CNN (Convolutional Neural Network) to estimate the blurring kernel and then applied a non-uniformsharpening model using prior knowledge of blocklevel images to remove image blur.Niu et al.[11] proposed a background difference method based on the Gaussian model, which updates the background in real time to avoid the influence of background interference on raindrop detection in medium rain, heavy rain and snow environments.Nah et al.[12]proposed end-to-end image sharpening based on deep learning, which uses amulti-scale convolutional neural network to directly restore the sharpness of blurred images.Kupyn et al.[13] used a feature pyramid model with generative adversarial networks to remove motion image blur, which provides a new conceptual approach to image sharpening.Tao et al.[14] proposed scaled recurrent networks,which introduces long and short memory networks to increase the acceptance domain and shared parameters to reduce the number of model parameters.Zhang et al.[15] used CNN to learn recurrent per location neural network (RNN) with pixel weights to increase the receptive domain.However,only a small portion of domain information can be captured with RNN or CNN and global contextdependent information cannot be obtained effectively.Zeng et al.[16] used a dense network for image sharpening, which can avoid the gradient disappearance problem, but the acceptance domain of the network is small and a limited amount of image information can be obtained.

    Wavelet transform [17,18] can divide the image signal into multi-directional subbands.Both wavelet transform and inverse transform can replace the down sampling and up sampling of CNN.Wavelet transform can describe the contextual dependencies and texture information of the image at different levels.High frequency subbands help to recover image texture details while constraining reconstruction of the low frequency subbands.Therefore, wavelet transform can be applied to image processing such as super resolution [19], image reconstruction [20], image defogging [21] and image sharpening [22], in which the combination of wavelet transform and deep learning is used for image sharpening.Wavelet transform images are used for decomposition extraction followed by deep learning to eliminate the data redundancy caused due to wavelet transform.

    3 The Proposed Model

    In this paper, we propose an improvedU-Net [23-25]model for biomassmaterial characterization image enhancement and the model structure is shown in Fig.1, which mainly consists of depthwise separable convolution [26], residual depthwise separable convolution [27], wavelet transform, a dense multiple receiver domain channel module (DMRFC) and a contextual channel attention module (CCA)..

    Figure 1: Model structure

    3.1 Model Structure

    3.1.1 Description of the Structure

    In the improvedU-Net network structure, the encoder is on the left and the decoder is on the right.The encoder implements the convolution operation with depthwise separable convolution instead of normal convolution and replaces the normal residual block with residual depthwise separable convolution.This aims to reduce the parameters in the network.Two-dimensional discrete wavelets are used instead of down sampling with the aim of obtaining contextual and texture information at different frequencies of the image, thereby reducing computational complexity and training difficulty.In deep coding, a DMRFC (dense multi-receptive field channel) module is used to obtain image information at different scales as well as to mitigate gradient disappearance and the emergence of reuse features.Decoding uses inverse wavelet permutation instead of the up sampling process to reduce the loss of image information.The CCA module is then used to fuse the splicing of image features from different network layers, which facilitates the propagation of detailed information from different layersand improves the sharpening effect.

    3.1.2 Loss Function

    The MSE (Mean Square Error) loss is often used as an evaluation metric for image sharpening by using the Euclidean distance to measure the difference between the predicted and actual values as shown in Eq.(1).However, the MSE loss function does not contain high-frequency information regarding the image, which may cause the model to be too smooth for the recovered image.Therefore,to obtain a clearer image, a perceptual loss function is used.This is expressed as the Euclidean distance between the sharpened imageRand the clear imageSin the convolution layer of VGGNet (Visual Geometry Group Network) as shown in Eq.(2).To further reduce the effect of noise in the sharpening process of the image as well as to recover the smoothness of the image space, the TV (Total Variation)loss function is used as shown in Eq.(3).Therefore, the loss function in this paperLtotalconsists of three components as shown in Eq.(4).

    In Eqs.(1)-(4),Rdenotes a sharpened image,Sdenotes a clear image, and ?(·) denotes the features obtained from the Conv 4-3 convolution layer in VGGNet.?xdenotes the gradient in the x-direction and?ydenotes the gradient in the y-direction.The parameters of the loss function and their values arew1=1,w2=0.05 andw3=0.001.

    3.2 Proposed Image Enhancement Method

    In this paper, we propose an improved U-Net network model for image enhancement and a sharpening method using the following approaches.

    (1) Incorporate the 2D DWT (Discrete Wavelet Transform) into the U-Net convolutional neural network.In the encoding stage, the original feature size is reduced and computational complexity is decreased.In the decoding stage, the sharpened image is recovered with the IWT(Inverse Wavelet Transform).

    (2) Introduce a DMRFC module in the U-Net network model.First, the potential features are extracted by improving the perceptual field through multiple receptive domains.The important channels and features are then selected by the channel attention mechanism to reduce the number of parameters.Finally, the potential feature information is reused via dense connectivity.

    (3) Improve the propagation of cross-layer information using the contextual channel attention module.The features between shallow and deep layers are fully fused to reduce the loss of image information and improve image reconstruction.

    3.2.1 Two-dimensional DWT

    In order to reduce the computational complexity and improve the quality of the reconstructed image, this paper mainly uses Haar wavelet transformation with two-dimensional DWT to divide the image signal into directional sub-bands so as to obtain different frequency information of the material characterization image.The one-dimensional high-pass filter is used φ(-).The low-pass filter ψ(-)is used to filter and vertically downsample each row of the image.Two filters are then used to filter and horizontally downsample each column.The wavelet transformation is represented as follows.

    Thexandyin Eqs.(5)-(8) denote thexandyaxes of the image information representing the image.IHLdenotes the horizontal high-frequency and vertical low-frequency information of the image representing the material.IHHdenotes the horizontal and vertical high-frequency information of the image representing the material.ILLdenotes the horizontal and vertical low-frequency information of the image representation the material.ILHrepresents the horizontal low-frequency and vertical highfrequency information of the image representing the material.In contrast, the wavelet inversion is the inversion of the four sub-images using the same filter.Upsampling ofIHL,IHH,ILHandILLis fused into the original image.Therefore, as shown in Fig.2, the original image is decomposed using DWT and then reconstructed using IWT without loss of information.A multi-level wavelet transform is performed and further processing ofIHL,IHH,ILH, andILLis carried out to obtain the transformation results of the multilevel wavelets.

    Figure 2: Two-dimensional discrete Haar wavelet transform with inverse transform

    3.2.2 DMRFC Module

    The DMRFC is shown in Fig.3 and the multi-receptive channel block is shown in Fig.4, which is a combination of the dilation acceptance domain block and the channel attention module.To increase the diversity of feature extraction in the dilation receptive domain, four feature extraction branches are used, using 3×3 convolution kernels with dilation rates of 1, 3, 5, and 7 respectively.The connection operation fuses the parallel feature maps of the four branches as shown in Eq.(9).The nonlinear features are adaptively weighted at each channel and the channel attention module in the attention mechanismmodule (CBAM) [28] is used for learning the weights of each channel.Merging the averag pooling features and maximum pooling features improves the nonlinear representation of the network and improves the sharpening ability using these two features.The output of the multi-acceptance channel block is shown in Eq.(10).

    Figure 4: Multi-receptive channel block

    In Eqs.(9), (10),xrepresents the input features, * is the convolution operation, andwis the convolution layer.The superscript represents the dilation convolution size and the subscript is the convolution kernelsize.LKrepresentstheLeaky ReLUactivationfunction,thecat[.]representsthe joint operation andCdenotes the fused features.The expressionMaxpoolin Eq.(10) represents the maximum pooling, theAvgpoolrepresents the average pooling andFCdenotes the fully connected layer.δis the Sigmoid activation function, andOutis the output.

    3.2.3 CCA Module

    In order to connect the low-level decoder and skip connection, obtain rich image semantic information and reduce information redundancy, the CCA [29] module is introduced as shown in Fig.5.The CCA feeds low-level image coding information (represented by LD) and skip connection information (represented by SC) into the CA(channel attention) module to obtain two types of feature information (represented by CA(LD) and CA(SC)) as shown in Eq.(11).The two are then summed as shown in Eq.(12) and the results obtained are each multiplied by the original information.Finally,they are fused as shown in Eq.(13).This method can fully fuse the information in LD and SC, which is beneficial for image sharpening.

    Figure 5: CCA block

    In Eqs.(11)-(13),GLPooldenotes global pooling, andConv1_BNdenotes 1×1 convolution followed by batch normalization.LKrepresents theLeaky ReLUactivation function,Conv1denotes 1×1 convolution andcat[.]represents the joint operation.

    4 Experimental Results

    4.1 Characterization Images Dataset

    In this paper, we used the biomass sample dataset obtained from the materials laboratory to train the model.In order to improve the generalization ability of the model,we performed data enhancement operations on the training set by randomly rotating as well as randomly flipping left, right, up and down.Gaussian noise was also added with rotation angles of 90, 180 and 270 degrees with a noise mean of 0 and variance of 0.0001.As a result, 7680 sets of images were obtained, with each set consisting of a blur and clear image pair.5760 sets of images were selected as the training set and 1920 sets of images were selected as the test set.

    4.2 Experimental Preparation

    To prevent overfitting of the model, the training set images were randomly cropped to a size of 128×128 pixels.The training period was set to 4000 rounds and the initial learning rate was set to 1e-4, which was halved every 1000 rounds.The network optimization method uses Adam [30], with parameters β1=0.9 and β2=0.999.The network model was built using the PyTorch deep learning framework and a GTX 2080 GPU.

    4.3 Quantitative Evaluation

    PSNR (Peak signal-to-noise ratio) and SSIM (structural similarity) are used as quantitative evaluation metrics, Larger values represent better image quality and they are calculated per Eqs.(14),(15).

    where MAX is the maximum value for the image (MAX=255) and RMSE is the root mean square error of the recovered clear and blurred images.μxandμyare the mean values of the imagexandyrespectively,,are the variance of imagexandyrespectively and σxyis the covariance ofxandy.c1andc2are small constants to avoid division by zero.

    Tab.1 shows the quantitative comparison of the average PSNR and average SSIM on the experimental test dataset with other methods.The comparison reveals that the method in this paper outperforms the other methods in terms of PSNR and SSIM.It is 0.42 dB higher than PSNR and 0.017dB higher than SSIM in the literature [14].

    Table 1: Quantitative evaluation of each algorithm on the dataset

    Tab.2 shows the time used on the experimental test dataset which is labeled as‘Time’.The model parameter size is labeled using‘Size’.The method used in this paper requires less time and smaller model parameters than those in the literature [12,14].

    Table 2: Running time and network model size for each algorithm on the dataset

    4.4 Characterization Image Enhancement Analysis

    Ten material characterization images were randomly taken from the test set and the average grayscale and gray gradient of the images before and after processing were analyzed as shown in Figs.6 and 7.Light gray represents the value of the image before processing and dark gray represents the value of the image after processing.From Fig.6, it can be seen that the average gray scale of the processed image has been improved.From Fig.7, it can be seen that the gray gradient of the processed image is lower than that of the unprocessed image, which helps to improve the quality of the material characterization image.

    Figure 6: Comparison of average greyscale before and after processing

    Figure 7: Comparison of gray gradient before and after processing

    To further verify the effectiveness of the method in this paper, a fuzzy biomass material characterization image (e.g., Fig.8a) was selected for more specific analysis as shown in Tab.3.From Tab.3, we can see that the original image has an average gray value of 138.5 and a gray gradient value of 14.2.Its average gray value is low and the gray gradient is too high, meaning that it is a lowquality fiber material characterization image which needs to be enhanced to better observe the surface morphology.After processing by the improved U-Net model in this paper, the enhanced sharpened image shown in Fig.8b is obtained.By visual observation, the sample brightness and contrast are improved, the image is smoother with less noise and the details are highlighted.The processed average gray values and gray gradient values are in the range of values for high-quality images when it comes to material characterization.The peak signal-to-noise ratio is 28.27 and a greater structural similarity quantitatively confirms the enhancement effect.

    4.5 Performance Impact Analysis of the Modules

    To verify the effectiveness of each module, three model experiments are conducted to illustrate the effect of dense multi-receiver domain channel blocks and CCA on the method in this paper.Model 1 contains only wavelet transformations, which replaces the dense multi-receiver domain blocks with five depth-separable residual blocks, with simple summation used instead of CCA.Model 2 contains wavelet transformations and dense multi-receiver domain channel blocks without CCA.Model 3 is the method used in this paper.The results are shown in Tab.4.

    Figure 8: Comparison of enhanced image sharpening effect

    Table 3: Analysis of quantitative parameter assessment

    Table 4: Quantitative results of different models

    From Tab.4, it can be seen that the Model 1 PSNR reaches 26.21dB after introducing wavelet transformation, which indicates that wavelet introduction can improve performance.This is due to the fact that wavelet transformation provides four types of sub-frequency information along with the potential to learn more detailed information.The Model 2 PSNR is slightly higher at 28.12dB, which indicates that the dense multi-receiver domain channel block helps to improve the quality of model recovery due to reduced information reuse, increased receiver domain and selection of important channel information.Model 3 introduces CCA, with the PSNR going up to 28.27dB in this paper,which indicates that CCA can further improve the model to obtain high-quality images.

    5 Conclusions

    An improved U-Net network model for material characterization image enhancement via image sharpening is proposed in this paper.It uses wavelet transformation to decompose and extract the frequency information of the blurred image and reduces the complexity of the computation.In order to extract the deep information in the image, a dense multi-receptor domain channel block is used to connect the multi-receptor domain channel blocks with dense connections.This reduces the multireceptor domain channel block parameters and enhances the transfer of features.The multi-receptor domain channel blocks can extract deep features and select important channels and features with the channel attention module.During the decoding process, the information between the skip connection and the low-resolution encoder is fused with the CCA module to produce a clearer image of the network.It has been shown experimentally that the method in this paper can significantly reduce the size of the model and reduces the computing time of the clear images recovered by the model.It also shows that the enhanced sharpened images have achieved better results.Information pickup from characteristic regions in materials science images plays a crucial role in helping R&D personnel in the preparation and characterization of new materials.

    The model proposed in this paper introduces more relevant modules and the next objective is to classify the samples into categories and conduct experiments for different modules to choose the optimal combination and to reduce the complexity of the model.

    Acknowledgement:The authors thanks Dr.Qiang Li for his help.The authors would like to thank the anonymous reviewers and the editor for the very instructive suggestions that led to themuch-improved quality of this paper.

    Funding Statement:This work was supported by the Fundamental Research Funds for Higher Education Institutions of Heilongjiang Province (135409505, 135509315, 135209245), the Heilongjiang Education Department Basic Scientific Research Business Research Innovation Platform“Scientific Research Project Funding of Qiqihar University”(135409421) and the Heilongjiang Province Higher Education Teaching Reform Project (SJGY20190710).

    Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.

    久久久国产成人免费| 舔av片在线| 俄罗斯特黄特色一大片| 最近在线观看免费完整版| 晚上一个人看的免费电影| 国产精品久久久久久av不卡| 精品人妻一区二区三区麻豆 | 免费黄网站久久成人精品| 99在线视频只有这里精品首页| 最近手机中文字幕大全| 日本欧美国产在线视频| 在线免费十八禁| 三级毛片av免费| 亚洲精品久久国产高清桃花| 国产色爽女视频免费观看| 人妻少妇偷人精品九色| 身体一侧抽搐| 在线天堂最新版资源| 国内精品美女久久久久久| 久久热精品热| 日韩欧美精品v在线| av国产免费在线观看| 国产精品日韩av在线免费观看| 九九久久精品国产亚洲av麻豆| 美女黄网站色视频| 男插女下体视频免费在线播放| 在线免费观看不下载黄p国产| 性色avwww在线观看| 国产成人91sexporn| 男人舔女人下体高潮全视频| av福利片在线观看| 日韩一区二区视频免费看| 精品久久久久久久人妻蜜臀av| 91狼人影院| 99视频精品全部免费 在线| 国内精品美女久久久久久| 日韩制服骚丝袜av| 亚洲精品久久国产高清桃花| 亚洲人成网站在线播| 日日摸夜夜添夜夜添av毛片| 午夜影院日韩av| 久久久久久国产a免费观看| 免费观看的影片在线观看| 欧美国产日韩亚洲一区| 在线观看66精品国产| 国产不卡一卡二| 欧美性感艳星| 成人二区视频| 国产精品国产高清国产av| 一a级毛片在线观看| 免费大片18禁| 国产单亲对白刺激| 欧美日韩国产亚洲二区| 日韩中字成人| 国产成人a∨麻豆精品| 老司机午夜福利在线观看视频| 亚洲中文日韩欧美视频| 2021天堂中文幕一二区在线观| 亚洲无线在线观看| 欧美xxxx性猛交bbbb| 国产大屁股一区二区在线视频| 精品福利观看| 国内少妇人妻偷人精品xxx网站| 日本 av在线| 亚洲av熟女| 如何舔出高潮| 网址你懂的国产日韩在线| 国产精品福利在线免费观看| 日日干狠狠操夜夜爽| 国产精品伦人一区二区| 精品福利观看| 黑人高潮一二区| 男插女下体视频免费在线播放| 亚洲天堂国产精品一区在线| 在线国产一区二区在线| 人妻久久中文字幕网| 久久久国产成人精品二区| 美女高潮的动态| 99热全是精品| 欧美色欧美亚洲另类二区| 午夜福利成人在线免费观看| 天堂影院成人在线观看| 久久久久久久久中文| 国产高清不卡午夜福利| 国产精品国产三级国产av玫瑰| 男插女下体视频免费在线播放| 真实男女啪啪啪动态图| 免费在线观看影片大全网站| 成人特级av手机在线观看| 午夜爱爱视频在线播放| 99久久精品一区二区三区| 日本免费一区二区三区高清不卡| or卡值多少钱| 亚洲欧美精品综合久久99| 久久鲁丝午夜福利片| 国产精品1区2区在线观看.| 91狼人影院| 免费看a级黄色片| 97在线视频观看| 少妇人妻精品综合一区二区 | 亚洲不卡免费看| 欧美激情国产日韩精品一区| 亚洲精品日韩av片在线观看| 69av精品久久久久久| 美女高潮的动态| 欧美日本视频| 国产乱人视频| 美女xxoo啪啪120秒动态图| 最近的中文字幕免费完整| 成人精品一区二区免费| 亚洲乱码一区二区免费版| 一级毛片久久久久久久久女| 人妻久久中文字幕网| 一个人看的www免费观看视频| 免费av毛片视频| 一级毛片我不卡| 久久久a久久爽久久v久久| 国产亚洲精品久久久久久毛片| 日韩欧美一区二区三区在线观看| 亚洲四区av| 欧美3d第一页| 天堂影院成人在线观看| av黄色大香蕉| 中文字幕免费在线视频6| 51国产日韩欧美| 久久精品夜色国产| 91精品国产九色| 亚洲人成网站高清观看| 欧美人与善性xxx| 身体一侧抽搐| 日本色播在线视频| 精品午夜福利在线看| 男女之事视频高清在线观看| 一级毛片我不卡| 色视频www国产| 国产在视频线在精品| 搡老熟女国产l中国老女人| 国语自产精品视频在线第100页| 最新中文字幕久久久久| 亚洲精品国产av成人精品 | 变态另类丝袜制服| 成年免费大片在线观看| 免费无遮挡裸体视频| 少妇被粗大猛烈的视频| 俺也久久电影网| 久久久精品大字幕| 国产三级在线视频| 在线观看午夜福利视频| 少妇熟女aⅴ在线视频| 精品一区二区三区人妻视频| 听说在线观看完整版免费高清| 男人狂女人下面高潮的视频| 国产熟女欧美一区二区| 成年女人永久免费观看视频| 成人亚洲精品av一区二区| 三级经典国产精品| 九色成人免费人妻av| 一进一出抽搐动态| 人妻制服诱惑在线中文字幕| 日韩三级伦理在线观看| 日韩 亚洲 欧美在线| 国产亚洲精品av在线| 97人妻精品一区二区三区麻豆| av视频在线观看入口| 99久国产av精品| 日日撸夜夜添| 久久久精品94久久精品| 日韩欧美精品免费久久| 欧美色欧美亚洲另类二区| 午夜精品一区二区三区免费看| 国产毛片a区久久久久| 日日啪夜夜撸| 久久久成人免费电影| 午夜精品一区二区三区免费看| 久久精品国产99精品国产亚洲性色| 99热精品在线国产| 欧美激情国产日韩精品一区| 小说图片视频综合网站| 97碰自拍视频| 国产一区二区在线观看日韩| 日本 av在线| 国产精品女同一区二区软件| 久久久色成人| 色av中文字幕| 日韩欧美在线乱码| 精品久久久久久久久亚洲| 国模一区二区三区四区视频| 国产精品野战在线观看| 婷婷色综合大香蕉| 在线免费十八禁| 国产伦一二天堂av在线观看| 少妇人妻一区二区三区视频| 丰满乱子伦码专区| 免费人成视频x8x8入口观看| 国产欧美日韩一区二区精品| 成人鲁丝片一二三区免费| 国产欧美日韩精品一区二区| 午夜a级毛片| 男人和女人高潮做爰伦理| 日韩在线高清观看一区二区三区| 最新中文字幕久久久久| 久久精品夜色国产| 日本黄色视频三级网站网址| 日本免费a在线| 国产aⅴ精品一区二区三区波| 日韩欧美 国产精品| 美女xxoo啪啪120秒动态图| 国产乱人偷精品视频| 欧美一区二区精品小视频在线| 日日干狠狠操夜夜爽| 丝袜喷水一区| 国产精品无大码| 男插女下体视频免费在线播放| 午夜免费男女啪啪视频观看 | 中出人妻视频一区二区| 不卡一级毛片| 国产黄色小视频在线观看| 在线播放国产精品三级| 神马国产精品三级电影在线观看| 国产成人影院久久av| 午夜视频国产福利| 久久精品久久久久久噜噜老黄 | 在线观看美女被高潮喷水网站| 亚洲第一电影网av| 亚洲精品一区av在线观看| 国产成人a∨麻豆精品| 亚州av有码| 在线免费观看不下载黄p国产| 人人妻人人澡人人爽人人夜夜 | 蜜桃亚洲精品一区二区三区| 麻豆久久精品国产亚洲av| 我要看日韩黄色一级片| 免费观看精品视频网站| 久久精品国产自在天天线| 毛片女人毛片| 白带黄色成豆腐渣| 我要看日韩黄色一级片| 直男gayav资源| 在线天堂最新版资源| 在线观看一区二区三区| 欧美激情在线99| 午夜精品国产一区二区电影 | 91久久精品国产一区二区三区| 免费观看人在逋| 亚洲人与动物交配视频| 亚洲av成人精品一区久久| 国产精品亚洲一级av第二区| 亚洲18禁久久av| 一边摸一边抽搐一进一小说| 国产av一区在线观看免费| 国产精品99久久久久久久久| 日韩欧美精品v在线| 亚洲av一区综合| 久久久久性生活片| 国产亚洲精品综合一区在线观看| 人人妻人人澡欧美一区二区| 国产精品女同一区二区软件| 久久草成人影院| 又黄又爽又免费观看的视频| 丰满的人妻完整版| 久久国内精品自在自线图片| 日韩欧美国产在线观看| 91久久精品电影网| 日本爱情动作片www.在线观看 | 一夜夜www| 亚洲丝袜综合中文字幕| 九九久久精品国产亚洲av麻豆| 国产亚洲欧美98| 人妻夜夜爽99麻豆av| 欧美激情在线99| 免费av观看视频| 3wmmmm亚洲av在线观看| 色播亚洲综合网| 好男人在线观看高清免费视频| 校园春色视频在线观看| 成人欧美大片| 亚洲激情五月婷婷啪啪| 综合色av麻豆| 久久久色成人| 亚洲丝袜综合中文字幕| 大型黄色视频在线免费观看| 欧美性猛交黑人性爽| 三级毛片av免费| 欧美日本视频| 国产 一区精品| 亚洲高清免费不卡视频| 亚洲精品国产成人久久av| 直男gayav资源| 美女大奶头视频| 久久久国产成人免费| av福利片在线观看| 久久综合国产亚洲精品| 国产精品人妻久久久久久| 国产精品人妻久久久影院| 色av中文字幕| 3wmmmm亚洲av在线观看| 国产中年淑女户外野战色| 日本黄色视频三级网站网址| 成人无遮挡网站| 亚洲国产精品合色在线| 午夜亚洲福利在线播放| 中文资源天堂在线| 久久久精品94久久精品| 国产在线男女| 国产精品三级大全| 国产精品电影一区二区三区| 亚洲国产欧洲综合997久久,| 精品一区二区三区av网在线观看| 91麻豆精品激情在线观看国产| 精品日产1卡2卡| 蜜臀久久99精品久久宅男| 简卡轻食公司| 身体一侧抽搐| 亚洲四区av| 在线观看免费视频日本深夜| 少妇猛男粗大的猛烈进出视频 | 波多野结衣高清无吗| 日本欧美国产在线视频| 成人亚洲欧美一区二区av| 69人妻影院| 国产亚洲精品av在线| АⅤ资源中文在线天堂| 亚洲欧美日韩高清在线视频| 色吧在线观看| 国产aⅴ精品一区二区三区波| 少妇丰满av| 久久草成人影院| 人人妻,人人澡人人爽秒播| 麻豆成人午夜福利视频| 国产成人精品久久久久久| 精品久久久久久久末码| av福利片在线观看| 嫩草影院入口| 日本熟妇午夜| 亚洲七黄色美女视频| 亚洲欧美成人精品一区二区| 哪里可以看免费的av片| 干丝袜人妻中文字幕| 中文在线观看免费www的网站| 真实男女啪啪啪动态图| 精品熟女少妇av免费看| 波多野结衣高清作品| 在线免费观看不下载黄p国产| 看免费成人av毛片| 91麻豆精品激情在线观看国产| 国产成人aa在线观看| 亚洲,欧美,日韩| 久久精品国产清高在天天线| 久久精品国产亚洲网站| 久久欧美精品欧美久久欧美| 日本撒尿小便嘘嘘汇集6| 久久久久久九九精品二区国产| 欧美色视频一区免费| 女同久久另类99精品国产91| 国产成人影院久久av| 波野结衣二区三区在线| 俺也久久电影网| 变态另类成人亚洲欧美熟女| 久久久久九九精品影院| 欧美三级亚洲精品| 精品一区二区三区视频在线| 国产成人精品久久久久久| 一本久久中文字幕| 三级国产精品欧美在线观看| 最近在线观看免费完整版| 麻豆久久精品国产亚洲av| 国产亚洲av嫩草精品影院| 国产亚洲精品久久久久久毛片| 精品久久久久久久末码| 成人特级黄色片久久久久久久| 六月丁香七月| 女同久久另类99精品国产91| 亚洲欧美清纯卡通| 国产免费一级a男人的天堂| 美女cb高潮喷水在线观看| 免费观看在线日韩| 97热精品久久久久久| .国产精品久久| 一区二区三区四区激情视频 | av在线蜜桃| 在现免费观看毛片| 色哟哟哟哟哟哟| 亚洲最大成人手机在线| 国产免费一级a男人的天堂| 99久国产av精品国产电影| 久久精品91蜜桃| 最近手机中文字幕大全| 国产免费一级a男人的天堂| 久久午夜亚洲精品久久| 成人av一区二区三区在线看| 精品人妻熟女av久视频| 欧美日本亚洲视频在线播放| a级毛色黄片| 欧美日韩在线观看h| 啦啦啦啦在线视频资源| 日韩中字成人| 亚洲美女搞黄在线观看 | 国产美女午夜福利| 男女那种视频在线观看| 免费黄网站久久成人精品| 国语自产精品视频在线第100页| 欧美性猛交╳xxx乱大交人| 午夜老司机福利剧场| 久久人妻av系列| 99久久久亚洲精品蜜臀av| 午夜老司机福利剧场| 91在线观看av| 亚洲熟妇中文字幕五十中出| 少妇人妻精品综合一区二区 | 国产成人影院久久av| 成人精品一区二区免费| 亚洲精品粉嫩美女一区| 久久精品国产亚洲av天美| 在线观看av片永久免费下载| 午夜老司机福利剧场| 国产亚洲精品av在线| 精品国产三级普通话版| 乱系列少妇在线播放| 国语自产精品视频在线第100页| 色哟哟·www| 最近最新中文字幕大全电影3| 国产乱人视频| 一区二区三区四区激情视频 | 亚洲天堂国产精品一区在线| ponron亚洲| 久久天躁狠狠躁夜夜2o2o| 欧美最黄视频在线播放免费| 久久久久久久午夜电影| 欧美+日韩+精品| 91在线精品国自产拍蜜月| 亚洲av二区三区四区| 一级毛片我不卡| 精品熟女少妇av免费看| 国产精品美女特级片免费视频播放器| 波多野结衣高清无吗| 性欧美人与动物交配| 3wmmmm亚洲av在线观看| 真实男女啪啪啪动态图| 精品久久国产蜜桃| 久久久久久久久久成人| 国产精品久久视频播放| 日本黄大片高清| 久久精品夜色国产| 少妇人妻一区二区三区视频| 国产乱人偷精品视频| 日韩三级伦理在线观看| 国产精品人妻久久久久久| 我要搜黄色片| 国产黄片美女视频| 毛片女人毛片| 日本五十路高清| 又爽又黄无遮挡网站| 国产精品av视频在线免费观看| 国内久久婷婷六月综合欲色啪| 国产一区亚洲一区在线观看| 看片在线看免费视频| 超碰av人人做人人爽久久| 精品一区二区三区视频在线观看免费| 美女cb高潮喷水在线观看| 国产女主播在线喷水免费视频网站 | 亚洲图色成人| 日本精品一区二区三区蜜桃| 91在线观看av| 免费人成视频x8x8入口观看| 国内精品美女久久久久久| 最近2019中文字幕mv第一页| 国产 一区精品| 亚洲国产精品久久男人天堂| 国产av不卡久久| 亚洲成av人片在线播放无| 听说在线观看完整版免费高清| 国产毛片a区久久久久| 97热精品久久久久久| 一区二区三区高清视频在线| 成年女人毛片免费观看观看9| 蜜臀久久99精品久久宅男| 精品午夜福利视频在线观看一区| 久久久成人免费电影| 99精品在免费线老司机午夜| 深爱激情五月婷婷| 免费av不卡在线播放| 在现免费观看毛片| 无遮挡黄片免费观看| 精品久久久久久久久久久久久| 国产男人的电影天堂91| 伦精品一区二区三区| 美女内射精品一级片tv| 国内精品美女久久久久久| 久久人人爽人人片av| 国内揄拍国产精品人妻在线| 亚洲欧美日韩高清在线视频| 日韩成人av中文字幕在线观看 | 国产淫片久久久久久久久| 久久久午夜欧美精品| 色av中文字幕| 国产男人的电影天堂91| 国产免费男女视频| 一边摸一边抽搐一进一小说| 国产女主播在线喷水免费视频网站 | 亚洲欧美精品自产自拍| 久久精品夜色国产| 最新中文字幕久久久久| 欧美最黄视频在线播放免费| 校园人妻丝袜中文字幕| 亚洲欧美精品自产自拍| 最新在线观看一区二区三区| 亚洲中文字幕日韩| 搡老妇女老女人老熟妇| 少妇的逼水好多| 国产黄色小视频在线观看| 国产熟女欧美一区二区| 99热全是精品| 亚洲综合色惰| 久久久久久久久中文| 99久久无色码亚洲精品果冻| 久久久久久九九精品二区国产| 99久国产av精品| 国产精品人妻久久久影院| 免费av观看视频| 亚洲五月天丁香| 成人毛片a级毛片在线播放| a级毛色黄片| 亚洲综合色惰| 国产精品不卡视频一区二区| 日韩欧美精品免费久久| 亚洲精品日韩在线中文字幕 | av黄色大香蕉| 91久久精品国产一区二区三区| 老熟妇乱子伦视频在线观看| 1024手机看黄色片| 精品久久久久久久久亚洲| 国产成人a区在线观看| 日本免费a在线| av.在线天堂| aaaaa片日本免费| 久久久成人免费电影| 午夜日韩欧美国产| 免费搜索国产男女视频| 老熟妇乱子伦视频在线观看| 国产精品久久久久久av不卡| 成年版毛片免费区| 简卡轻食公司| 九九久久精品国产亚洲av麻豆| 亚洲人成网站在线观看播放| 六月丁香七月| 日产精品乱码卡一卡2卡三| 午夜久久久久精精品| 狂野欧美激情性xxxx在线观看| 日本成人三级电影网站| 在线看三级毛片| 国产精品国产三级国产av玫瑰| 国产成人一区二区在线| 禁无遮挡网站| 91av网一区二区| 青春草视频在线免费观看| 日本黄色片子视频| 免费不卡的大黄色大毛片视频在线观看 | 蜜臀久久99精品久久宅男| 日本在线视频免费播放| 亚洲精品国产成人久久av| 在线观看66精品国产| 99国产精品一区二区蜜桃av| 不卡视频在线观看欧美| 综合色av麻豆| 免费一级毛片在线播放高清视频| 在线免费观看的www视频| 成年女人毛片免费观看观看9| 亚洲精品日韩av片在线观看| 极品教师在线视频| 婷婷精品国产亚洲av在线| 日本欧美国产在线视频| 国产精品1区2区在线观看.| 少妇被粗大猛烈的视频| 老司机影院成人| 国产精品亚洲一级av第二区| 国产精品三级大全| 亚洲精品乱码久久久v下载方式| 又粗又爽又猛毛片免费看| 久久久色成人| 国产免费一级a男人的天堂| 亚洲欧美日韩高清专用| 久久久久久九九精品二区国产| 老司机影院成人| 老熟妇仑乱视频hdxx| 久久午夜亚洲精品久久| 22中文网久久字幕| 精品一区二区三区视频在线| 日本黄大片高清| 春色校园在线视频观看| 亚洲国产高清在线一区二区三| 欧美成人一区二区免费高清观看| 精品福利观看| 免费看美女性在线毛片视频| 日韩在线高清观看一区二区三区| a级毛片免费高清观看在线播放| 亚洲人成网站在线观看播放| 色综合站精品国产| 天堂av国产一区二区熟女人妻| 国产av不卡久久| 成人综合一区亚洲| 亚洲内射少妇av| 日韩欧美国产在线观看| 亚洲一级一片aⅴ在线观看| 久久久a久久爽久久v久久| 日韩欧美三级三区| 成人美女网站在线观看视频| 国产亚洲91精品色在线| 免费一级毛片在线播放高清视频| 91久久精品国产一区二区三区| 99视频精品全部免费 在线| 极品教师在线视频| 91久久精品国产一区二区三区| 亚洲第一区二区三区不卡| 成年av动漫网址| 日韩欧美免费精品|