• <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 ?

    Parkinson’s Detection Using RNN-Graph-LSTM with Optimization Based on Speech Signals

    2022-08-24 12:59:12AhmedAlmasoudTaiseerAbdallaElfadilEisaFahdAlWesabiAbubakarElsafiMesferAlDuhayyimIshfaqYaseenManarAhmedHamzaandAbdelwahedMotwakel
    Computers Materials&Continua 2022年7期

    Ahmed S.Almasoud, Taiseer Abdalla Elfadil Eisa, Fahd N.Al-Wesabi, Abubakar Elsafi,Mesfer Al Duhayyim, Ishfaq Yaseen, Manar Ahmed Hamza,*and Abdelwahed Motwakel

    1Department of Information Systems, College of Computer and Information Sciences, Prince Sultan University, Mahayil,62529, Saudi Arabia

    2Department of Information Systems-Girls Section, King Khalid University, Mahayil, 62529, Saudi Arabia

    3Department of Computer Science, College of Science and Artsat Mahayil, King Khalid University, Mahayil, 62529,Saudi Arabia

    4Faculty of Computer and IT, Sana’a University, Sana’a, Yemen

    5Department of Software Engineering, College of Computer Science and Engineering, University of Jeddah, Saudi Arabia

    6Department of Natural and Applied Sciences, College of Community-Aflaj, Prince Sattam bin Abdulaziz University,Mahayil, 62529, Saudi Arabia

    7Department of Computer and Self Development, Preparatory Year Deanship, Prince Sattam bin Abdulaziz University,AlKharj, 62529, Saudi Arabia

    Abstract: Early detection of Parkinson’s Disease (PD) using the PD patients’voice changes would avoid the intervention before the identification of physical symptoms.Various machine learning algorithms were developed to detect PD detection.Nevertheless, these ML methods are lack in generalization and reduced classification performance due to subject overlap.To overcome these issues, this proposed work apply graph long short term memory (GLSTM)model to classify the dynamic features of the PD patient speech signal.The proposed classification model has been further improved by implementing the recurrent neural network (RNN) in batch normalization layer of GLSTM and optimized with adaptive moment estimation (ADAM) on network hidden(ADAM)on network hidden layer.To consider the importance of feature engineering, this proposed system use Linear Discriminant analysis (LDA) for dimensionality reduction and Sparse Auto-Encoder(SAE)for extracting the dynamic speech features.Based on the computation of energy content transited from unvoiced to voice(onset) and voice to voiceless (offset), dynamic features are measured.The PD datasets is evaluated under 10 fold cross validation without sample overlap.The proposed smart PD detection method called RNN-GLSTM-ADAM is numerically experimented with persistent phonations in terms of accuracy,sensitivity, and specificity and Matthew correlation coefficient.The evaluated result of RNN-GLSTM-ADAM extremely improves the PD detection accuracy than static feature based conventional ML and DL approaches.

    Keywords: Dimensionality reduction; LDA; feature extraction; auto encoder;classification; recurrent neural network; LSTM and optimization

    1 Introduction

    Parkinson‘s disease (PD) is a neurodegenerative disorder which affects the human brain nerve cells.This is considered to be the second common disease after Alzheimer disease [1] that targets the peoples age more than 60 years.Most of the PD patients are diagnosed in the age of 70 and also 15% of cases are occurred under the age of 50.PD symptoms can be characterized as motor and non motor symptoms.PD motor symptoms includes tremor, stiffness, slowness of movement and postural instability.PD non motor symptoms includes cognitive dysfunction, mood disorders, sensory dysfunction, pain and dysautonomia [2].Among these two motor symptoms are very common in PD patients.PD detection based on phonation is the proven method with 90% effectiveness [3]which shows voice impairments.Hence, Voice impairments through speech signals are considered to be the one of the earlier symptom thatmight not be noticeable to the listeners.PDvocal dysfunction is characterized as such as reduced tongue flexibility, longer pauses, reduced vocal volume, narrowing and variations in pitch range, articulation rate and voice intensity level.Acoustic analysis is considered as non-invasive tool by many researches to diagnose PD.Further more, early diagnosis of PD is done with phonation and speech data by many researchers [4-6].Diagnosing the Parkinson’s disease is difficult.It diagnosed through brain scans and clinical assessments which are very cost, leads to erroneous sometime and need some professional monitoring.

    Speech signal processing is becomes the interdisciplinary research that includes artificial intelligence and signal processing.In order to solve human-computer interaction (HCI) issues, more methods are developed by many researchers [7].Based on this many Machine Learning (ML)algorithms are applied to detect PD patients [8-15].ML algorithms are considered PD detection as a classification task.Machine learning is a scheme that analyze the data, automatically learn the system data and attitudes [16].ML can be divided into traditional ML algorithm and Deep Learning (DL)based algorithms.DLis inspired bybrain signal processingthatused programmed neural network[17]to make accurate decisions by the machine without human interference.

    ML methods such as Artificial Neural Network [18], K-Nearest Neighbors (KNN) [19], Extreme Gradient Boosting (XGBoost) [20], Random Forest (RF) [21], Support Vector Machine (SVM),Decision Tree (DT) [22], Genetic algorithm (GA) [23] were used for PD classification using speech signals.

    Neural network model with more generalization provides the accurate result on the analysis while the data tested before in the training also used for testing.With recurrent neural network (RNN) with Long Short Term Memory (LSTM), the input data are learned which consists of chunks of memory to retain the input information [24,25].This paper illustrates how the classifiers variation, parameters representation and feature extraction are affect the performance of the model with more accuracy,efficient and robust.The major contribution of the paper is as follows:

    ?Proposed a PD detection model using static and dynamic features using deep learning model for earlier diagnosis.To scale the input features values in a range, preprocessing method called min-max normalization used.

    ?To create most relevant low dimensional space, Linear Discriminant Analysis based dimensionalityreduction method used which will transform the data into low dimensional space execution

    ?Proposed a classification model called RNN-Graph LSTM to use the dynamic features of speech signals for PD detection.For each batch normalization layer, the proposed RNNGLSTM has been applied after hidden layers to standardize the hidden layer output.

    ?In order to improve the classification model performance, ADAM optimizer used to adjust the network weight.Neural network optimization is used to increase the accuracy and also reduces the loss.

    ?The proposed models with speech signal dataset is experimented and quantitative results are compared with conventional PD detection model.The proposed model have generalization abilities that are not identified by previous studies.This model also be used for early diagnosis of PD prediction with lower complexity.

    The rest of the paper is as follows: Section 2 discusses about the review of the literature, Section 3 stated about the dataset used for evaluation, Section 4 proposed an efficient PD classification model,Section 5 discusses about the experimented results and Section 6 concludes the work with future directions.

    2 Related Work

    In order to differentiate PD patients from healthy people, ML and DL techniques would be the best tool.This section reviews various ML and DL based PD prediction system from speech signals.Sharma et al.[26] proposed a method to extract the voice signal features such as MFCC, Jitter,Shimmer, glottal pulse and pitch.Sakar et al., [27] proposed Tunable Q-Factor wavelet transform(TQWT) method to predict PD patients using voice signals.This dataset was experimented using DL methods called convolutional neural networks [28].

    Bouchikhi et al., [29] proposed relief-F feature selection model with SVM classifier.This method chosen 10 features from 22 features.SVM classifier with 10 fold cross validation proves that Relief-F feature selection showed 96.88% accuracy.Experimented dataset consists of 195 voice samples.Relief based algorithms (RBA) are susceptible to noise intrusive with nearest neighbor.Hemmerling et al.[30]proposed nonlinear SVM with PCA based PD detection.This model obtained 93.43% of accuracy on PD classification.The experimented dataset was small and lack in lower precision in prediction.

    Parisi et al.[31] proposed multilayer perceptron with lagrangian SVM based classifier to detect PD patients.The relevant features were assigned by custom cost functions by MLP which consists of both accuracy and AUC score.MLP extract 20 most important and relevant features with the score value.This proposed model obtained 100% accuracy than compared algorithms.Hybrid method for preprocessing and PD classification.Subtractive clustering features weighting (SCFW) was proposed as a preprocessing method to reduce the dataset variance.They proposed kernel based extreme learning machine (KELM) as a classifier and justified the efficiency of KELM in terms accuracy, sensitivity and specificity, Kappa statistic value and Roc curve.

    Caliskan et al., [32] proposed DNN classifier for PD detection that consists of stacked auto encoder to extract the vocal features.This work is compared with traditional ML models and concludes that DNN based PD prediction secure high accuracy.Moreover, DNN needs more data in training phase and also it took more training time for parameter space search.

    In [33] CNN has been used to extract the speech features from short time fourier transform and wavelet transform.Main issue of CNN is to model long distance contextual data while using dilated convolution layers [34].Recurrent neural network (RNN) overcome this issue and able to model long distance contextual data by saving the previous computations.However, RNN based method suffer from gradient problem and it is hard to tune the parameters in the network layer.LSTM [35] has been used to overcome this issue.

    In this paper RNN trained LSTM with graph structure has been proposed as a classification model for PD detection.The classification accuracy is further improved by ADAM optimizer based on various speech features.RNN-GLSTM can hope up with larger dataset without increase the model size.Compared to traditional time series models, GLSTM have more effective as it learn long term dependencies that follow the previous time proceeding and forward to next layers.This proposed model also acquire additional benefit on the use of preprocessing and feature extraction methods.It overcomes the disadvantages of previous methods such as limited dataset size and features that reduce the accuracy for PD prediction due to traditional NN use feed forward layer.RNN-GLSTM used loop network in forward and backward that enhance the accuracy of PD prediction.

    3 Data Set

    The data set used for this analysis is created by Max Little of the University of Oxford which is the collaboration with National centre for voice and speech [36], Denver.The speech signals are recorded by Colorado.The details of the dataset are contained in Tab.1.This data set consists of 31 peoples biomedical voice measurements.Among them 23 patients are with Parkinson’s disease.The column of the table is voice measure of particular person.Each row of the table corresponds to one of the 195 voice recordings of the individuals.Based on the status column in the table with binary value, aim is to classify the unhealthy (PD patients with value 1) from healthy persons (value 0).This data set consists of 24 attributes which includes number of frequencies (low, medium, high),number of variations in terms of frequency called Jitter and its types such as MDVP:Jitter(%),MDVP:Jitter(Abs), MDVP:RAP, MDVP:PPQ, Jitter:DDP [37,38] and number of variations in terms of amplitude called shimmer and its types such as MDVP:Shimmer, MDVP:Shimmer(dB), Shimmer:APQ3, Shimmer:APQ5, MDVP:APQ, Shimmer:DDA.MDVP:Fo(hz)-average vocal fundamental frequency, MDVP:Fhi(Hz)-average vocal fundamental frequency and MDVP:Flo(Hz)-maximum vocal fundamental frequency, Three nonlinear fundamental frequency variations such as Spread1,Spread2 and PPE, two measures of ratio of noise such as NHR and HNR.It is an unbalanced dataset.Sample data from the dataset is shown in Fig.1.

    Table 1: Dataset description

    Figure 1: Sample data

    4 Proposed Model Description

    The model for proposed PD diagnostic system is shown in Fig.2.The voice speech signals are preprocessed with normalization to remove the null values.This model uses training database to construct the model.The generalization of the proposed model has been tested through testing database.The classification system performance is further improved with dimensionality reduction and feature extraction techniques.Dimensionality reduction is the mechanism to reduce the high dimensional data space into low dimensional space.Feature extraction is the procedure to select relevant features from feature set to improve the accuracy of the classifier [39,40].There are numerous dimensionality reduction and feature extraction methods are available.In this work, we use Linear Discriminant Analysis(LDA)for dimensionality reduction and Sparse Auto encoder (SAE) for feature extraction.Each of these methods is explained in upcoming sections.

    4.1 Data Pre-processing

    Raw data are inconsistent due to it contains lots of error and null values.These raw data are transformed into understandable format through pre processing phase that will improve the results.Good preprocessing method yields good classification result.The min-max normalization process is denoted in Eq.(1) for the data set D.After normalization the values are lie in the range [0,1].

    4.2 Training Testing Data Model

    The entire dataset is divided into three parts such as training dataset and testing dataset with the ration of 80:20 respectively.Based on the cross validation, the training set is further divided into training and validation dataset in the ratio 80:20 respectively.Fig.3 shows the cross validation sets which the classification performed.

    4.3 Dimensionality Reduction Using LDA

    Various feature extraction and classification approaches are used original pattern vector with low dimensionality.To meet this objective, this proposed work use Linear Discriminant Analysis (LDA)for dimensionality reduction method at initial stage of classification model for this, LDA uses Fisher ratio which is denoted as in Eq.(2)

    Figure 2: Overview of proposed PD diagnostic model

    Figure 3: Cross validation ratio of dataset

    where,σ1and σ2-variances of first and second classes respectively and μ1-μ2-difference between the mean values of two classes.The LDA maximizes the Fisher ratio by maximizes the scatter between two classes called Sinterand reduces the variance that minimizes the scatter among the class called Sintra.The fisher ration of Eq.(1) is rewritten as,

    The main objective of phase is to transform the data into lower dimensional space by maximize the Eq.(3).To consider this, LDA use transformation matrix called w withSinterandSintraas in Eqs.(4)and (5).

    Hence Eq.(3) becomes

    Transformation matrix w is evaluated by the calculation of eigenvectors of.Thus LDA has been used to transform the p dimensional data into k dimension data where k<=(n-1), n is the number of classes in the dataset.In this proposed work, considered dataset is a binary class dataset.So n=2 and k=1 (healthy or PD patient).With the maximization of class separability of Eq.(2),

    4.4 Feature Extraction Using SAE

    Sparse Autoencoder (SAE) is an unsupervised deep neural network based feature extraction method with single hidden layer [41] to encodes the given data.This also estimates the error and eextracts the relevant features through hidden layer expressions [42].Autoencoder have some functional flaws.AE cannot find features through inputting and copying memory into its implicit layer [43].To overcome this issue, sparsity based auto encoder is proposed [44].SAE or sparsity regularization is denoted in Eq.(7).

    where KL-Kullback-Leibler (KL) divergence,-activation function of jth hidden node,ρ-sparsity parameter.KL is mathematically denoted as in Eq.(8)

    SAE has been trained with the cost function stated in Eq.(9) which consists of three terms such as Mean Squared Error (MSE) denoted in Eq.(10) which reconstruct input X intoover entire training dataset [45], second is lasso regression term denoted in Eq.(11) and third term is sparsity transformation stated in Eq.(7).

    where N-total number of input data points, X-input,-reconstructed output,α-coefficient of lasso regression and β-coefficient of sparsity regularization.Lasso regression is used with SAE to compress the least important features coefficient into zero to denote it as not relevant for further process.This will shrink the parameter space.Mathematical representation of Lasso regression denoted in Eq.(11).

    where l=layer,nl=number of layers,ul-number of units in the layer l and-weight value between ithnode of layer l and jthnode of layer l+1.Lasso regression can add the magnitude absolute value as penalty term.To improve the feature extraction process, this penalty coefficient is set as zero.SAE with three hidden layers is shown in Fig.4 which consists of three parts such as encoder, compression or representation of latent space and decoder.

    Figure 4: SAE with three hidden layers

    Through hidden layers, encoder encodes the input data to latent space and decoder decodes the latent space code to output layer.Sigmoid activation function is used for nonlinear mapping.Rectified Linear Unit (ReLU) is not suitable for auto-encoder due to the capability of dealing negative value as zero.This will reduce network training ability.The considered dataset feature values also consist of negative values.Hence, Sigmoid activation function is used rather than ReLU.Objective of SAE is to learn the speech signal features from the input data.

    4.5 Classification Using RNN-GLSTM

    Recurrent Neural Network (RNN) is the generalized form of feed forward neural network with internal memory.The output of RNN relies with previous computation and sent back to the recurrent network.Internal memory in the RNN is used to operate the input series and make the decision.Long short term memory (LSTM) is based on back propagation for training.LSTM consists of three gates such as input gate, forget and output gate.Input gate used sigmoid activation function which is used to decide the values of input that modify the memory.Forget gate is responsible to decide the details that are to be discarded from previous state and output gate responsible to control the output.Compare to traditional LSTM, in graph LSTM each tree node represents the single LSTM unit, Fig.5.

    Figure 5: Proposed RNN-LSTM classification model

    This model consists of seven layers such as input layer, five hidden layers and an output layer.Recurrent Neural network is comprised with the input layer of LSTM cell.Each input layer of LSTM layer represents the Phonation Features (PF) of the speech signals.23 features are represented by 23 neurons in the input layer of LSTM cell which is shown in Fig.6.

    where, b-bias vector, w-weight matrices and H-hidden layer function of each feature.

    Figure 6: GLSTM cell

    For the forward GLSTM, the parameters in Fig.6 such as input gate it, forget gate ft, output gate otand cell state ctfor particular iteration t with activation function σ are updated as in following Eqs.(15) from to (19)

    For backward GLSTM, the parameters are updated as follows in Eqs.(20) to (24).

    4.6 Optimization of RNN-GLSTM with ADAM

    The main objective of machine and deep learning is to reduce the difference between actual and predicted output which is called as cost or loss function.In order to choose the optimal value for the cost function, weight in cost function is updated using optimization algorithm while training the neural network.This will leads to enhance the result of prediction model.This proposed work uses ADAM optimizer [46,47] to update the weight.It is the most suggested optimization method for deep learning networks [48,49].It takes advantages of stochastic gradient descent (SGD) algorithm and root mean square (RMS).This proposed PD classification model chose ADAM optimization due to the advantages such as it does not consume large memory, it uses second moment called uncentred variance of the gradients and first moment called mean.

    Optimization algorithm

    Step 1: Initialize 1stmoment m1=0 and second moment m2=0 and iteration t=0

    Step2: bias of 1stand 2ndmoment are updated according to Eqs.(20) and (21)

    Step 3: bias corrected of 1stand 2ndmoment are calculated according to Eqs.(22) and (23)

    Step 4: parameters of RNN-GLSTM is updated as in Eqs.(24) to (27)

    where β1and β2-hyper parameters (values are 0.8 and 0.88 respectively),ε-learning rate (10-2).It is implemented in python scikit-learn library.Thus the way the training data of proposed model has been optimized.Next, testing data are applied on this trained optimized RNN-GLSTM for classification.

    5 Proposed System Evaluation and Discussions

    This section discusses about the validation and evaluation metrics used and evaluated results.Performance of the proposed system is compared with existing ML algorithms to prove the efficiency of the proposed PD diagnostic model.For experiment, dataset used in Section 3 was used and compared with conventional ML algorithms such as Multi layer perceptron (MLP), K-Nearest Neighbor (KNN), Random forest (RF) and Principal component analysis (PCA) with Support Vector Machine (SVM) are exhibited.

    5.1 Evaluation Metrics

    The output for the proposed RNN-GLSTMmodel has been evaluated with the evaluation metrics such as accuracy, precision, recall, F1 score and Mathew’s correlation Coefficient (MCC) based on confusion matrix shown in Tab.2, Fit time and score time .

    Table 2: Confusion matrix-binary classification

    From Tab.2, accuracy, precision, recall, F1 score and MCC are denoted in Eq.from (28) to (32).MCC can be used with imbalanced dataset due to its robust evaluation characteristics.

    Fit time:It is the time to fit the estimator on training data set for each cross validation split.

    Score time:It is the time for scoring the estimator on testing dataset for each cross validation split.

    5.2 Experimental Results

    The proposed PD diagnostic model RNN-GLSTM optimized by ADAM has been implemented in python Scikit-learn 0.22.1.The parameter settings of traditional ML models are shown in Tab.3.

    Table 3: ML algorithms parameter settings

    The simulation result of the proposed system with dimensionality reduction and optimization are shown in Tab.3.For evaluation, various number of hyper parameters (H1, H2 and H3) are used for each layer.In our proposed work, 3 hidden layers are used.H1 is the number of neurons in first hidden layer, H2 is the number of neurons in second hidden layer and H3 is the number of neurons in third hidden layer of RNN-LSTM-ADAM in Tab.4.

    Table 4: Evaluation of proposed system with hyper parameters

    5.3 Comparative Analysis

    Experimental result in terms of different input speech signal on ML algorithms with proposed is discussed.

    The MLP, RF and PCA-SVM were implemented in order to evaluate the performance of the proposed model.With the same set of training and testing data, PD prediction is made for all the models.From Fig.7,it can be observed that random forest obtained least efficiency than other models and proposed RNN-GLSTM obtained high efficiency than other models in terms of the evaluation metrics.The proposed model obtained high accuracy of 91.21%, high F1 score of 85.02%, high precision of 86.72%, high recall value of 85.9% and high MCC value of 0.352.

    Figure 7: Performance evaluation of proposed vs. conventional methods

    Comparison in terms of standard dimensionality reduction and feature extraction methods and optimization.Compared to standard PCA with proposed classification model, proposed LDA with classification model obtained 93.5% of accuracy.Compared to standard ICA based feature extraction with proposed model shown in Tab.5.

    Table 5: Different evaluation of proposed system

    6 Conclusion

    This paper presented an efficient PD diagnose model called Recurrent Neural Network based graph LSTM optimized with Adam optimizer.This proposed model used min-max normalization to scale the data values in the range [0,1], LDA based dimensionality reduction to transform the dataset into low dimensional space, SAE based feature extraction to consider the most relevant features for classification.This work used PD speech signal dataset and considered the dynamic features.Multiple evaluation in terms of dimensionality reduction, feature extraction methods were computed and the models is evaluated with various evaluation metrics.This classification model obtained 95.4% of accuracy, 93.4% of F1 score and 0.865 of MCC.In future, more features such as handwriting features are also considered for classification and explore this model to applicable on multi label classification with more deep learning based architectures.

    Acknowledgement:The authors would like to acknowledge the support of Prince Sultan University for paying the Article Processing Charges (APC) of this publication.

    Funding Statement:The authors extend their appreciation to the Deanship of Scientific Research at King Khalid University for funding this work under Grant Number (RGP 1/282/42).www.kku.edu.sa.

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

    狂野欧美白嫩少妇大欣赏| 国产毛片在线视频| 日本欧美国产在线视频| av免费在线看不卡| 国产高清不卡午夜福利| 亚洲不卡免费看| 中文乱码字字幕精品一区二区三区| av又黄又爽大尺度在线免费看| 亚洲成人av在线免费| 国产精品久久久久久久久免| 久久99热这里只频精品6学生| 大话2 男鬼变身卡| h视频一区二区三区| 欧美 日韩 精品 国产| 日日爽夜夜爽网站| 一级毛片 在线播放| 色网站视频免费| 91成人精品电影| 国模一区二区三区四区视频| 亚洲国产av影院在线观看| 午夜福利在线观看免费完整高清在| 欧美精品高潮呻吟av久久| 少妇猛男粗大的猛烈进出视频| 久久久久久伊人网av| 成人亚洲欧美一区二区av| 精品亚洲成a人片在线观看| 亚洲高清免费不卡视频| 精品久久久久久久久亚洲| 日韩,欧美,国产一区二区三区| 99热网站在线观看| 亚洲av综合色区一区| 亚洲欧洲日产国产| 亚洲第一区二区三区不卡| 国产色爽女视频免费观看| 精品一区在线观看国产| 亚洲精品自拍成人| 在线观看免费视频网站a站| 99久久精品国产国产毛片| 久久久国产一区二区| 免费高清在线观看视频在线观看| 亚洲国产日韩一区二区| 特大巨黑吊av在线直播| 久久午夜福利片| 亚洲精品色激情综合| 男的添女的下面高潮视频| 亚洲精华国产精华液的使用体验| 高清毛片免费看| 狂野欧美白嫩少妇大欣赏| 高清不卡的av网站| 少妇被粗大猛烈的视频| 两个人的视频大全免费| 久久久久国产网址| 午夜免费男女啪啪视频观看| 另类精品久久| 亚洲高清免费不卡视频| 男女边吃奶边做爰视频| 日韩精品有码人妻一区| 亚洲精品日韩在线中文字幕| 国产一区二区三区综合在线观看 | 精品卡一卡二卡四卡免费| 一边亲一边摸免费视频| 在线观看免费日韩欧美大片 | 免费久久久久久久精品成人欧美视频 | 热99国产精品久久久久久7| 在线观看三级黄色| 成人手机av| 肉色欧美久久久久久久蜜桃| 亚洲av成人精品一区久久| 午夜精品国产一区二区电影| 伦理电影免费视频| av播播在线观看一区| 22中文网久久字幕| 日本黄色日本黄色录像| 亚洲人成网站在线播| 欧美日韩国产mv在线观看视频| 一级片'在线观看视频| 国产黄频视频在线观看| 国产高清不卡午夜福利| 精品视频人人做人人爽| 亚洲国产精品成人久久小说| 黑人欧美特级aaaaaa片| 高清av免费在线| 欧美人与性动交α欧美精品济南到 | 国产无遮挡羞羞视频在线观看| 天堂8中文在线网| 26uuu在线亚洲综合色| av.在线天堂| 亚洲第一av免费看| 三上悠亚av全集在线观看| 性色avwww在线观看| 国产国语露脸激情在线看| 97在线视频观看| 一区二区三区精品91| 久久精品国产亚洲av涩爱| 成人国语在线视频| 欧美三级亚洲精品| 免费看不卡的av| 久久久久久久久久久丰满| 99热6这里只有精品| 啦啦啦视频在线资源免费观看| 国产精品免费大片| 国产精品无大码| 欧美xxⅹ黑人| 亚洲无线观看免费| 亚洲av成人精品一二三区| 国产女主播在线喷水免费视频网站| 国产精品.久久久| 欧美激情 高清一区二区三区| h视频一区二区三区| 久久久久久久国产电影| 婷婷成人精品国产| 国产成人aa在线观看| 校园人妻丝袜中文字幕| 中文字幕最新亚洲高清| 黑人高潮一二区| 久久精品久久久久久噜噜老黄| 久久久亚洲精品成人影院| 亚洲av二区三区四区| 熟女av电影| 18+在线观看网站| 久久精品熟女亚洲av麻豆精品| av在线老鸭窝| 欧美一级a爱片免费观看看| 欧美一级a爱片免费观看看| 青青草视频在线视频观看| 3wmmmm亚洲av在线观看| 国产精品成人在线| 夜夜看夜夜爽夜夜摸| kizo精华| 特大巨黑吊av在线直播| av天堂久久9| 大陆偷拍与自拍| 亚洲国产精品999| 国产精品蜜桃在线观看| 女性被躁到高潮视频| 人人妻人人添人人爽欧美一区卜| 国产高清国产精品国产三级| 看十八女毛片水多多多| 一区二区日韩欧美中文字幕 | 久久久精品94久久精品| 97超视频在线观看视频| 蜜臀久久99精品久久宅男| 国产熟女欧美一区二区| 精品人妻熟女毛片av久久网站| 日韩在线高清观看一区二区三区| 在线观看www视频免费| 久久精品久久精品一区二区三区| 亚洲综合色惰| 黄片无遮挡物在线观看| 欧美xxⅹ黑人| .国产精品久久| 精品视频人人做人人爽| 精品亚洲成a人片在线观看| 丰满饥渴人妻一区二区三| 水蜜桃什么品种好| av电影中文网址| 美女国产视频在线观看| 亚洲色图 男人天堂 中文字幕 | 亚洲欧美清纯卡通| av天堂久久9| 亚洲情色 制服丝袜| 亚洲综合色网址| 桃花免费在线播放| 国产极品粉嫩免费观看在线 | 国产欧美另类精品又又久久亚洲欧美| 欧美精品国产亚洲| 日本wwww免费看| 如日韩欧美国产精品一区二区三区 | 黑丝袜美女国产一区| 亚洲,欧美,日韩| 国产精品99久久久久久久久| 精品少妇黑人巨大在线播放| 免费大片黄手机在线观看| 国产精品一区二区三区四区免费观看| 精品少妇黑人巨大在线播放| 欧美人与善性xxx| 人妻制服诱惑在线中文字幕| 国产成人精品福利久久| 亚洲欧美成人精品一区二区| 国产在线视频一区二区| 精品国产露脸久久av麻豆| 在线看a的网站| 你懂的网址亚洲精品在线观看| 亚洲第一区二区三区不卡| 欧美精品人与动牲交sv欧美| 亚洲精品乱码久久久久久按摩| 中文天堂在线官网| 男女边摸边吃奶| 日本色播在线视频| 日韩中字成人| 中文精品一卡2卡3卡4更新| 精品99又大又爽又粗少妇毛片| 婷婷色av中文字幕| 啦啦啦啦在线视频资源| 丰满饥渴人妻一区二区三| 美女视频免费永久观看网站| 18禁观看日本| 亚洲美女搞黄在线观看| 伊人久久国产一区二区| 欧美精品高潮呻吟av久久| 蜜桃在线观看..| 久久毛片免费看一区二区三区| 人妻一区二区av| 成人亚洲精品一区在线观看| 免费av不卡在线播放| 日韩熟女老妇一区二区性免费视频| 国产免费现黄频在线看| 欧美日韩一区二区视频在线观看视频在线| 欧美精品一区二区大全| 欧美日韩在线观看h| 在线观看免费日韩欧美大片 | 美女cb高潮喷水在线观看| a 毛片基地| 丝瓜视频免费看黄片| 免费黄网站久久成人精品| 亚洲图色成人| 黑人巨大精品欧美一区二区蜜桃 | 中文乱码字字幕精品一区二区三区| 曰老女人黄片| 熟女人妻精品中文字幕| 亚洲av不卡在线观看| 午夜福利视频在线观看免费| 国产精品人妻久久久久久| 99国产综合亚洲精品| 久久免费观看电影| 91aial.com中文字幕在线观看| 91国产中文字幕| 少妇被粗大猛烈的视频| 男人操女人黄网站| 欧美日韩视频精品一区| 国产一区二区三区av在线| 国产午夜精品久久久久久一区二区三区| 亚洲一区二区三区欧美精品| 两个人的视频大全免费| 亚洲熟女精品中文字幕| 国产午夜精品久久久久久一区二区三区| 亚洲精品,欧美精品| 观看av在线不卡| 黄片无遮挡物在线观看| 亚洲美女黄色视频免费看| 久久99精品国语久久久| 黄片无遮挡物在线观看| 满18在线观看网站| 成人毛片60女人毛片免费| 日本黄大片高清| 三级国产精品欧美在线观看| 成人毛片a级毛片在线播放| 亚洲精品aⅴ在线观看| 九草在线视频观看| 另类亚洲欧美激情| 久久婷婷青草| 欧美精品高潮呻吟av久久| 国产成人午夜福利电影在线观看| 丝袜喷水一区| 亚洲国产毛片av蜜桃av| 国产淫语在线视频| 精品一区二区三卡| 18+在线观看网站| 性高湖久久久久久久久免费观看| 最近中文字幕2019免费版| 国产熟女欧美一区二区| 久久久亚洲精品成人影院| 最新的欧美精品一区二区| 91aial.com中文字幕在线观看| 亚洲综合色惰| 亚洲一级一片aⅴ在线观看| 国国产精品蜜臀av免费| 久久精品久久精品一区二区三区| 亚洲四区av| 99热这里只有精品一区| 久久久欧美国产精品| 国产有黄有色有爽视频| 亚洲美女黄色视频免费看| 国产欧美另类精品又又久久亚洲欧美| 观看美女的网站| 免费日韩欧美在线观看| 飞空精品影院首页| 国产成人精品一,二区| 国产探花极品一区二区| 永久网站在线| 国产一区二区在线观看日韩| 91久久精品国产一区二区三区| 亚洲欧美日韩卡通动漫| 午夜激情福利司机影院| 国产亚洲av片在线观看秒播厂| 午夜福利,免费看| 欧美日韩一区二区视频在线观看视频在线| 69精品国产乱码久久久| 久久人人爽人人爽人人片va| 亚洲婷婷狠狠爱综合网| 99热这里只有是精品在线观看| tube8黄色片| 考比视频在线观看| 十八禁网站网址无遮挡| 久久精品国产鲁丝片午夜精品| 纯流量卡能插随身wifi吗| 九九爱精品视频在线观看| 久久精品国产亚洲av天美| 美女中出高潮动态图| 欧美日本中文国产一区发布| 欧美xxⅹ黑人| 天天操日日干夜夜撸| 日本与韩国留学比较| 最新中文字幕久久久久| 中文字幕最新亚洲高清| 久久综合国产亚洲精品| 久久国内精品自在自线图片| 在线免费观看不下载黄p国产| 久久精品国产a三级三级三级| 国产淫语在线视频| 成人综合一区亚洲| 老司机亚洲免费影院| 精品久久久久久久久av| 99re6热这里在线精品视频| √禁漫天堂资源中文www| 永久免费av网站大全| 亚洲精品aⅴ在线观看| 日韩av不卡免费在线播放| 亚洲伊人久久精品综合| 少妇的逼好多水| 观看美女的网站| 少妇猛男粗大的猛烈进出视频| 午夜激情av网站| 亚洲五月色婷婷综合| 99久久人妻综合| 欧美激情国产日韩精品一区| 亚洲精品aⅴ在线观看| 久久久久精品久久久久真实原创| 日韩av免费高清视频| 免费观看无遮挡的男女| 99热全是精品| 美女中出高潮动态图| 99精国产麻豆久久婷婷| 街头女战士在线观看网站| 色婷婷av一区二区三区视频| 色视频在线一区二区三区| 久久久久精品性色| 亚洲精品乱码久久久v下载方式| 精品人妻熟女毛片av久久网站| 天天躁夜夜躁狠狠久久av| 国产精品三级大全| 性色avwww在线观看| 97精品久久久久久久久久精品| 国产精品麻豆人妻色哟哟久久| 日韩熟女老妇一区二区性免费视频| 国产精品一国产av| 国产精品三级大全| 亚洲精品av麻豆狂野| 亚洲精品456在线播放app| 精品视频人人做人人爽| 成年美女黄网站色视频大全免费 | 久久精品夜色国产| 成人综合一区亚洲| 国产熟女欧美一区二区| 人人妻人人澡人人看| 在线观看美女被高潮喷水网站| 三级国产精品片| 国产午夜精品一二区理论片| a级毛片在线看网站| av又黄又爽大尺度在线免费看| 卡戴珊不雅视频在线播放| 免费黄网站久久成人精品| 国产一区二区三区综合在线观看 | 人妻夜夜爽99麻豆av| 亚洲精品色激情综合| 一本一本综合久久| 老女人水多毛片| 99精国产麻豆久久婷婷| 成人国语在线视频| 国模一区二区三区四区视频| 99久久人妻综合| 你懂的网址亚洲精品在线观看| 亚洲精品色激情综合| 国产片特级美女逼逼视频| 免费人成在线观看视频色| 国产永久视频网站| 2018国产大陆天天弄谢| 黄色视频在线播放观看不卡| av专区在线播放| 男女啪啪激烈高潮av片| 国产精品欧美亚洲77777| 国产亚洲精品久久久com| 精品一区二区三卡| 视频在线观看一区二区三区| 在线观看国产h片| 妹子高潮喷水视频| videosex国产| 九九久久精品国产亚洲av麻豆| 国产在视频线精品| 青春草视频在线免费观看| 精品人妻偷拍中文字幕| 一级a做视频免费观看| a级毛片黄视频| 久热这里只有精品99| 街头女战士在线观看网站| 亚洲欧美日韩卡通动漫| 涩涩av久久男人的天堂| 国产免费福利视频在线观看| 777米奇影视久久| 人妻少妇偷人精品九色| 少妇的逼水好多| 91久久精品电影网| 亚洲国产欧美在线一区| 亚洲欧美一区二区三区国产| 人妻一区二区av| 国产欧美日韩综合在线一区二区| 大片免费播放器 马上看| 亚洲欧美色中文字幕在线| 亚洲精品视频女| 天美传媒精品一区二区| 日日啪夜夜爽| 亚洲av日韩在线播放| 搡老乐熟女国产| 久久亚洲国产成人精品v| 国产精品国产三级专区第一集| 18禁在线播放成人免费| 国产乱人偷精品视频| 国产女主播在线喷水免费视频网站| 国产片内射在线| 黑人巨大精品欧美一区二区蜜桃 | 欧美最新免费一区二区三区| 亚洲av福利一区| 熟女人妻精品中文字幕| 亚洲国产日韩一区二区| av又黄又爽大尺度在线免费看| 国产精品一区www在线观看| 中国美白少妇内射xxxbb| 少妇人妻久久综合中文| 最近的中文字幕免费完整| 亚洲精品美女久久av网站| 18禁裸乳无遮挡动漫免费视频| 水蜜桃什么品种好| 99国产精品免费福利视频| 国产成人免费观看mmmm| 97精品久久久久久久久久精品| www.av在线官网国产| 日韩三级伦理在线观看| 在线观看www视频免费| 亚洲成色77777| 丰满乱子伦码专区| 午夜精品国产一区二区电影| 成人毛片60女人毛片免费| av在线播放精品| tube8黄色片| 亚洲熟女精品中文字幕| 一本大道久久a久久精品| 亚洲久久久国产精品| 日韩三级伦理在线观看| 丰满饥渴人妻一区二区三| 国产午夜精品久久久久久一区二区三区| 亚洲国产精品一区三区| 九九爱精品视频在线观看| 亚洲精品aⅴ在线观看| 日本av手机在线免费观看| 亚洲av在线观看美女高潮| 日韩av在线免费看完整版不卡| 最近中文字幕高清免费大全6| 亚洲国产毛片av蜜桃av| 黑丝袜美女国产一区| 一级毛片电影观看| 九九爱精品视频在线观看| 欧美日韩在线观看h| 下体分泌物呈黄色| 亚洲在久久综合| 亚洲一级一片aⅴ在线观看| 男人爽女人下面视频在线观看| 欧美变态另类bdsm刘玥| 3wmmmm亚洲av在线观看| 亚洲欧美清纯卡通| 国产精品不卡视频一区二区| 久久精品国产自在天天线| 99久国产av精品国产电影| 久久毛片免费看一区二区三区| 欧美日韩精品成人综合77777| av电影中文网址| 亚洲国产精品999| 99热网站在线观看| 久久精品国产亚洲网站| 波野结衣二区三区在线| √禁漫天堂资源中文www| 欧美成人精品欧美一级黄| 午夜福利在线观看免费完整高清在| 青青草视频在线视频观看| 日韩精品免费视频一区二区三区 | 麻豆成人av视频| 99热网站在线观看| 另类精品久久| 亚洲经典国产精华液单| 国产精品欧美亚洲77777| 亚洲精品乱码久久久v下载方式| 亚洲国产精品一区二区三区在线| 一区二区三区四区激情视频| 亚洲精品久久成人aⅴ小说 | 久久午夜综合久久蜜桃| 国产精品麻豆人妻色哟哟久久| 亚洲美女黄色视频免费看| av女优亚洲男人天堂| 日韩熟女老妇一区二区性免费视频| 丰满乱子伦码专区| tube8黄色片| 欧美bdsm另类| 建设人人有责人人尽责人人享有的| 纵有疾风起免费观看全集完整版| 嫩草影院入口| 日韩在线高清观看一区二区三区| 美女国产视频在线观看| 妹子高潮喷水视频| 国产永久视频网站| 在线观看免费高清a一片| 欧美老熟妇乱子伦牲交| 亚洲国产欧美在线一区| 超碰97精品在线观看| 久久久久久久久久久久大奶| 亚洲欧美清纯卡通| 五月开心婷婷网| 一级毛片电影观看| 亚洲精品中文字幕在线视频| 久久久午夜欧美精品| 9色porny在线观看| 精品一区二区三卡| 一本—道久久a久久精品蜜桃钙片| 亚洲成人一二三区av| freevideosex欧美| 男人爽女人下面视频在线观看| 最近中文字幕2019免费版| 亚洲,一卡二卡三卡| 国产毛片在线视频| 97在线视频观看| 久久国产精品男人的天堂亚洲 | 日本与韩国留学比较| 欧美精品亚洲一区二区| 你懂的网址亚洲精品在线观看| 国产精品久久久久久av不卡| 99re6热这里在线精品视频| 成人漫画全彩无遮挡| av专区在线播放| 五月玫瑰六月丁香| 高清在线视频一区二区三区| 成人黄色视频免费在线看| 最近2019中文字幕mv第一页| 日本-黄色视频高清免费观看| 精品国产国语对白av| 国产精品一区二区三区四区免费观看| 一级二级三级毛片免费看| 欧美变态另类bdsm刘玥| 2018国产大陆天天弄谢| 免费黄频网站在线观看国产| 亚洲精品日韩av片在线观看| 欧美丝袜亚洲另类| 一区二区三区精品91| 99热全是精品| 亚洲国产精品999| 亚洲欧洲日产国产| 亚洲第一av免费看| 亚洲av.av天堂| 蜜桃久久精品国产亚洲av| 黄片无遮挡物在线观看| 最近的中文字幕免费完整| 欧美精品一区二区大全| 在现免费观看毛片| 插逼视频在线观看| 91精品伊人久久大香线蕉| 精品少妇黑人巨大在线播放| 久久综合国产亚洲精品| 国产精品不卡视频一区二区| 日韩av免费高清视频| 久久婷婷青草| 高清av免费在线| 精品一品国产午夜福利视频| 一边亲一边摸免费视频| 久久精品熟女亚洲av麻豆精品| 久久人人爽人人爽人人片va| 精品久久久噜噜| 久久人人爽人人爽人人片va| 亚洲第一av免费看| 高清av免费在线| 亚洲精品久久成人aⅴ小说 | 亚洲综合精品二区| 狠狠婷婷综合久久久久久88av| 99久久精品国产国产毛片| 久久人妻熟女aⅴ| 免费日韩欧美在线观看| 交换朋友夫妻互换小说| 欧美日韩av久久| 婷婷色麻豆天堂久久| 99热这里只有是精品在线观看| 国产黄频视频在线观看| 飞空精品影院首页| 精品久久久久久久久亚洲| 亚洲精品中文字幕在线视频| 一区在线观看完整版| 中文字幕免费在线视频6| 日本wwww免费看| 少妇精品久久久久久久| 国产成人精品婷婷| 午夜视频国产福利| 青春草亚洲视频在线观看| 99re6热这里在线精品视频| 国产av国产精品国产| 波野结衣二区三区在线| av电影中文网址| 欧美精品一区二区大全| 51国产日韩欧美| 在线观看免费日韩欧美大片 | av在线观看视频网站免费| 另类精品久久| 国产精品秋霞免费鲁丝片| 一个人免费看片子| 国产国拍精品亚洲av在线观看| 国产av一区二区精品久久| 成人漫画全彩无遮挡| 九草在线视频观看| 一区二区三区乱码不卡18| 亚洲经典国产精华液单| 亚洲综合精品二区|