• <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 ?

    AnimeNet:A Deep Learning Approach for Detecting Violence and Eroticism in Animated Content

    2023-12-12 15:50:46YixinTang
    Computers Materials&Continua 2023年10期

    Yixin Tang

    Department of Cooperative Course of Performance,Film&Animation,Sejong University,Seoul,05006,Korea

    ABSTRACT Cartoons serve as significant sources of entertainment for children and adolescents.However,numerous animated videos contain unsuitable content,such as violence,eroticism,abuse,and vehicular accidents.Current content detection methods rely on manual inspection,which is resource-intensive,time-consuming,and not always reliable.Therefore,more efficient detection methods are necessary to safeguard young viewers.This paper addresses this significant problem by proposing a novel deep learning-based system,AnimeNet,designed to detect varying degrees of violent and erotic content in videos.AnimeNet utilizes a novel Convolutional Neural Network(CNN)model to extract image features effectively,classifying violent and erotic scenes in videos and images.The novelty of the work lies in the introduction of a novel channel-spatial attention module,enhancing the feature extraction performance of the CNN model,an advancement over previous efforts in the literature.To validate the approach,I compared AnimeNet with state-of-the-art classification methods,including ResNet,RegNet,ConvNext,ViT,and MobileNet.These were used to identify violent and erotic scenes within specific video frames.The results showed that AnimeNet outperformed these models,proving it to be well-suited for real-time applications in videos or images.This work presents a significant leap forward in automatic content detection in animation,offering a high-accuracy solution that is less resource-intensive and more reliable than current methods.The proposed approach enables it possible to better protect young audiences from exposure to unsuitable content,underlining its importance and potential for broad social impact.

    KEYWORDS Computer vision;animation;deep learning;classification;attention mechanism

    1 Introduction

    1.1 Motivation

    It is well-documented that seeing violence and explicit content has a negative impact on society[1].Children’s thoughts may be influenced in a favorable or unfavorable way by the animation films on the multimedia platform[2].While these materials may be used to educate children,they also have the capacity to entirely captivate children and teenagers,leading to less time spent interacting with others.Despite the fact that most children’s content is good,entertaining,or educational,recent research has highlighted the development of dangerous content[3,4].A well-known instance of this phenomenon is the Elsagate incident[5,6],in which con artists combined popular animated characters like Spiderman,Disney’s Frozen,Mickey Mouse,and others with unpleasant features like light violence and explicit behavior.According to research,children who often watch violent cartoons and movies are less sensitive to pain and more prone to act aggressively [7].These films cause them to start behaving aggressively at home and at school.For example,bullying is a widespread problem,especially in schools,and it has a devastating effect on everyone involved.Moreover,the increase of explicit scenes in animation content also affects the well-being of small children.

    The term “explicit scene” refers to anything,such as nudity and sex,that is objectionable,inexcusable,or goes against socially acceptable norms.As a result,material made for kids has to be given specific attention.The current filtering procedure for an explicit scene involves manually screening the whole piece of material.Furthermore,some explicit scenes are merely romantic scene that has been considered reasonable for children to watch and falls under the‘U’(unrestricted)category[8].Therefore,the animation industry does not ban certain explicit scenes from animation content.

    Despite existing research on the impact of explicit content,there’s a lack of comprehensive mechanisms for detecting such content in animations.The current procedure for filtering explicit scenes relies heavily on manual screening.While some explicit scenes are marked as‘U’(unrestricted)because they’re considered reasonable for children [9],the industry lacks a rigorous,standardized method to filter out inappropriate scenes.

    There is a crucial need for automated deep learning-based classification systems to identify and filter explicit or violent content in animation films.This will help to protect children from exposure to inappropriate content and its potential adverse effects.Automated monitoring should be designed to accurately detect violent or explicit scenes and categorize them based on their level of explicitness.The monitoring system may take benefit from a comprehensive database of animation content marked with their explicit levels to train the deep learning models.

    1.2 Related Work

    A wide variety of methodologies have been proposed to enable automatic violence or explicit scene detection in cartoon images,each with its own advantages,disadvantages,and areas of application.In general,these methods can be grouped into two categories:violent content detection and pornographic content detection.

    In terms of violent content detection,Ditsanthia et al.[10] and Sumon et al.[11] both utilized a combination of deep convolutional networks and long short-term memory (LSTM) for violence detection.While these works achieved impressive accuracies,they were both limited by their lack of consideration for the degree of violence and specificity of real-world violence.Chen et al.[12]extended these models by integrating 3D convolutional neural networks(CNN)and support vector machines(SVM),though at a high computational cost that hindered real-time applications.Ye et al.[13]took a different approach by focusing on feature extraction methods,yet their work also did not extend to animated or cartoon videos.

    Two works stand out for their explicit attention to cartoon violence.Yousaf et al.[14]proposed an EfficientNet-BiLSTM method for detecting and classifying inappropriate cartoon content.However,they only classified full frames,thereby potentially discarding useful information from noninappropriate portions of the frame.Meanwhile,Khan et al.[15] developed a MobileNet-based violence detection model for cartoons,which despite achieving high accuracy,showed overfitting issues,limiting its real-world application.

    For pornographic content detection,Wang et al.[16] and Mohamed [17] both proposed CNNbased models,although their approaches face difficulties with real-time detection and robust classification,respectively.Perez et al.[18]and Yuan et al.[19]developed models that achieved high accuracies on their respective datasets,but the former’s model required the fusion of multiple CNNs and an SVM,and the latter’s dataset was not substantial enough for comprehensive classification.Finally,Shen et al.[20]presented an ensemble framework with an uncertain evidence-based method,though they,like all others in this category,did not account for cartoon and animation videos.The detailed content is shown in Table 1.

    Table 1:Comparative overview of various machine learning and deep learning methods applied for content detection and classification,highlighting their specific goals,key contributions,and notable limitations

    Existing literature on the automatic detection of violent or pornographic content in images presents several areas of improvement.These include the incorporation of degrees of violence or explicitness in the classification process,the adaptation of these techniques specifically for animation and cartoon content,the optimization of model complexity to ensure real-time application feasibility,and a comprehensive violence and pornographic dataset for training deep learning model to detect both scenes simultaneously.

    Addressing these identified gaps,my study focuses on developing an automated,novel deep learning-based classification system that can classify both violent and pornographic scenes.The proposed system is designed to accurately detect and classify violent or explicit scenes in cartoon and animation videos.Crucially,it considers the degree of explicitness and violence,thus providing a nuanced and more thorough analysis compared to existing methods.Concurrently,it is designed with a balance between computational complexity and performance,ensuring that real-time detection and classification are feasible.This study also presents a comprehensive dataset for researchers to further process and improve the detection of such scenes.

    1.3 Contribution

    This study presents a novel CNN-based system for the automatic detection and classification of violence and erotic levels in animated and cartoon images.The primary contributions of this paper include:

    1.The creation of a novel dataset comprising 4,044 high-resolution (1920 × 1080) images featuring violent and erotic content derived from animated and cartoon sources.This dataset serves as a valuable resource for developing and testing classification models in this domain.

    2.The development of AnimeNet,an innovative CNN-based model designed to effectively classify violence and erotic levels in animated content.This novel model demonstrates superior performance in identifying and categorizing various degrees of unsuitable content within animated images.

    3.The introduction of a cutting-edge Channel-Spatial attention module significantly improves feature extraction capabilities for animated and cartoon images.This module enhances the overall effectiveness of the AnimeNet model,ensuring accurate and reliable classification results.

    The rest of the paper is structured as follows:Section 2 presents the methods used in this study,Sections 3 and 4 discusse and analyze the extensive experiments to evaluate the performance of the proposed model,and the conclusion is drawn in Section 5.

    2 Erotic and Violence Level Classification System

    Computer vision is a rapidly developing field that has the potential to revolutionize a wide range of industries,including robotics,education,healthcare,and biology.Deep learning-based computer vision solutions,such as object detection and classification,have become increasingly important in various applications.

    In this context,I propose an erotic and violence level classification system using deep learning models,which has a well-defined architecture,as illustrated in Fig.1.The system is composed of four major components,namely Image Collection,Data Augmentation,Model Training,and Erotic and Violence Level Classification Output.

    To create the training dataset,I retrieved images from animation series and movies and examined them manually to identify those containing pornographic and violent content.These images were classified into seven classes based on the severity of the content.Next,I fine-tuned the training dataset using various data augmentation techniques to improve the model’s generalization capability.

    Several deep learning models were trained on the augmented dataset to evaluate their performance in classifying pornographic and violence levels.Finally,I selected the best-performing model to classify images into the seven predefined classes.

    2.1 Dataset Preparation

    The identification of erotic or violent content in animation is a crucial challenge that requires a systematic and reliable approach.In this study,I identified 30 animation series and movies that feature a higher proportion of such scenes than other well-known animation content.

    To extract relevant frames from the videos,a python code was written automatically processed each frame.Human evaluators were then employed to select the extracted frame which contained violence or pornographic content.The videos ranged in duration from 40 min to 1 h,and the resolution of all frames was up to 1980×1080.

    From this process,I extracted a total of 15000 frames containing pornographic or violent content.I then selected 3890 images with distinct features of pornographic or violent content to construct my dataset.To facilitate classification,I further categorized the dataset into seven classes,including three levels of eroticism(Levels 1,2,and 3),three levels of violence(Levels 1,2,and 3),and a normal class.The class distribution of the dataset is shown in Table 2.

    Table 2:The description of the proposed pornographic and violence level classification dataset

    Overall,my approach offers a systematic and rigorous methodology for identifying and extracting pornographic or violent content from animation.The resulting dataset can be used for a wide range of applications,including developing machine learning models for content classification and filtering.

    2.2 Data Augmentation

    Data augmentation is a widely adopted technique for enhancing the quality of a dataset and mitigating overfitting challenges.Datasets with distinct characteristics require specific data augmentation approaches to yield optimal performance.One such method proposed by Cubuk et al.is the AutoAugment technique,which automatically selects the most suitable augmentation method for each image in a dataset[21].

    The AutoAugment approach utilizes a policy-based methodology,where each policy comprises various sub-policies that apply augmentation techniques to images based on their individual characteristics.

    In prior studies,AutoAugment was evaluated on the CIFAR-10 dataset,achieving a low error rate of 1.48 percent.Furthermore,using the ImageNet dataset,AutoAugment attained a Top-1 accuracy of 83.54 percent.The policy derived from the ImageNet dataset was then deployed to other datasets,including the Stanford Cars and FGVC Aircraft image identification tasks,where it significantly reduced the error rate by 1.16 percent and 1.76 percent,respectively.

    In this study,the proposed dataset comprises images with varying degrees of violence and complex characteristics that require appropriate augmentation techniques to enhance the dataset’s quality.To achieve this,I have utilized the AutoAugment approach to automatically process the images from the dataset and improve their quality before training the proposed model.In Fig.2,three sample images are presented,showcasing the application of the AutoAugment method on the proposed dataset.The algorithm intelligently selects and applies horizontal flipping,brightness enhancement,and rotation transformations based on the data-driven augmentation policy it has discovered.

    Figure 2:Autoaugment method applied to the dataset images,showcasing the policy-based selected transformations.Original image,(a)flip,(b)brightness enhancement,(c)rotation

    2.3 AnimeNet Model

    This study proposes the use of AnimeNet,a deep learning model,to extract features and classify the level of erotic and violent content in detected frames,as illustrated in Fig.3.The AnimeNet model consists of three main parts:stem,body,and head.

    The body of the AnimeNet model comprises four stages,with each stage containing severaldiblocks.The detailed structure of the stem,stage,and head can be seen in Fig.4.Furthermore,the structure of individualdiblocks is depicted in Fig.5.

    The stem component of the AnimeNet model serves as the input and performs initial feature extraction,followed by the body that further extracts relevant features using thediblocks.Finally,the head component performs the classification task to determine the level of pornographic or violent content in the detected frame.

    Figure 3:The structure of the proposed AnimeNet model

    Figure 4:The description of the proposed AnimeNet model includes stem,stage,down sampling,and head parts

    Figure 5:The structure of the individual block,which is used in every stage of the body parts of the proposed AnimeNet

    The proposed AnimeNet model’s architecture is designed to enhance its performance and robustness in handling sensitive content classification tasks.Its multi-stage design allows for deep feature extraction and enables the model to learn and represent complex image features effectively.Additionally,the use ofdiblocks enhances the model’s ability to learn and identify low-level features while maintaining computational efficiency.

    I developed a novel CNN-based deep learning model called AnimeNet,as illustrated in Fig.4,to achieve superior performance in handling complex violence and erotic content classification tasks.The proposed model comprises several key components,including stem,stage,downsampling,and head parts.

    The stage component of the proposed model contains a series of blocks that are responsible for learning and representing complex image features.Stages 1,2,and 4 consist of three blocks each,while stage 3 contains nine blocks.The detailed structure of the blocks can be seen in Fig.5.

    The stem component of the proposed model comprises a 4×4 convolutional layer with a stride of 4,followed by batch normalization and the Exponential Linear Unit (ELU) activation function.The ELU activation function is known to be effective in reducing the vanishing gradient problem and enhancing the performance of deep learning models.

    The downsampling component of the proposed model comprises a 2×2 max pooling layer with a stride of 2,a 3×3 convolutional layer with a stride of 1 and padding,and a batch normalization layer.These layers are responsible for downsampling the input image and extracting relevant features.

    The head component of the proposed model includes an average pooling layer followed by fully connected layers,which are responsible for classifying the frame into different categories based on the level of violence and pornographic content.

    Fig.5 provides an illustration of the individual blocks in every stage of the body parts of the AnimeNet model.Each block follows a specific design and sequence of layers optimized for handling complex image classification tasks effectively.

    The first layer in each block is a 1×1 convolutional layer followed by batch normalization and the Exponential Linear Unit(ELU)activation function.This initial layer is responsible for reducing the dimensionality of the input image while preserving its key features.

    Next,a 3×3 convolutional layer is used to increase the channel size of the input image by four times.This layer is also followed by batch normalization and the ELU activation function,which enhance the model’s ability to learn and represent complex features effectively.

    An enhanced attention module is then added to extract important features from the input image and improve the model’s performance.This module helps the model to focus on the most relevant regions in the input image,enhancing its ability to detect and classify violence and pornographic content accurately.

    Finally,a 1 × 1 convolutional layer is added to decrease the channel size to its previous dimensionality.This layer is also followed by batch normalization and the ELU activation function,which help to maintain the consistency of the model’s features throughout the network.The relevant equation for the individual block can be defined as follows,

    Overall,the design of the individual blocks in the AnimeNet model is optimized for achieving superior performance in handling complex violence and pornographic content classification tasks.The incorporation of specific layers and modules,such as the enhanced attention module,enhances the model’s ability to learn and represent complex image features while maintaining computational efficiency.

    Fig.6 depicts the enhanced attention module used in the proposed AnimeNet model,which comprises both spatial and channel attention networks.The design of this module is inspired by study[22],which has been shown to be effective in improving the performance of deep learning models for image classification tasks.

    Figure 6:The overall structure of the proposed channel-spatial attention module

    The channel attention blocks in the enhanced attention module recalibrate the channels of the input image by incorporating global spatial information.To achieve this,adaptive average pooling layers are used,which provide a receptive field of the whole spatial extent at each stage of the model,aiding content classification.The processed features are then passed through a 1×1 convolution layer,followed by a sigmoid activation function.The sigmoid function maps the input values to a range of 0 to 1,making it useful for gating or scaling the features.The relevant equation for channel attention block can be defined as follows:

    The spatial attention blocks in the enhanced attention module contain a 1×1 convolution layer followed by a sigmoid activation function.The element-wise max operation is used to perform an element-wise competitiveness between the two attention blocks,providing selective spatial and channel excitation.The relevant equation for spatial attention block can be defined as follows,

    The output from EAM ensures that the final classification is based on the most relevant features in the input image,improving the overall accuracy of the model.The relevant equation can be defined as follows:

    The enhanced attention module in the AnimeNet model is a key component that enables the model to selectively focus on the most relevant features in the input image,enhancing its ability to detect and classify violence and pornographic content accurately.By incorporating both spatial and channel attention networks,the enhanced attention module provides a powerful mechanism for learning and representing complex image features.

    Fig.7 presents the overall flow chart of the proposed animation content classification system using the AnimeNet model.The proposed system starts with an input in the form of a video.This video is processed through a specialized capture program designed to extract individual frames,resulting in an amassed collection of images.These images form the basis for the subsequent steps and are pivotal for the overall classification task.

    Figure 7:Flowchart depicting the process of animation video content classification using the AnimeNet model

    The image collection then undergoes an‘Auto Augment’process.This particular phase involves the augmentation of the original images through various transformations such as rotation,scaling,shifting,and flipping.The objective here is to increase the diversity and volume of training data,thereby enhancing the robustness of the model and its ability to generalize.

    Once the images have been suitably augmented,they are subject to an additional round of processing.During this ‘Processing Images’phase,images are manipulated to meet specific model requirements.This could entail operations such as resizing to maintain uniformity,normalization for optimal model performance,or conversion into the necessary formats or color spaces.

    The processed images are subsequently deployed for model training.This stage entails the employment of the conditioned images to train the deep learning model,enabling it to recognize and differentiate between diverse classes present in the animation frames.

    Upon successful completion of the training phase,a ‘Trained Model’emerges.This model,enriched by the insights gleaned from the training data,is now equipped to predict the class of new,unseen animation images.

    The proposed system also encompasses a decision-making procedure.If an image under examination is an animation image,the model proceeds to predict its class.In contrast,if the image does not qualify as an animation image,it is categorized as‘No animation image.’Finally,the output of the system is in the form of classified frames derived from the original video input.

    3 Experiment and Result Analysis

    3.1 Experimental Parameters

    In this research,I chose Ubuntu as the experimental platform.The Intel(R)Core(TM)i7-8700@3.20 GHz processor and the NVIDIA GeForce GTX TITAN X GPU were chosen as the platform for this research,while Python 3.7 and Pytorch 1.11 were chosen as the programming language and deep learning framework,respectively.The detailed hyperparameter and experimental settings used in this study can be seen in Table 3.

    Table 3:Key parameters and specifications employed for the experimental setup in the current study

    An extensive experiment was done on the proposed AnimeNet model to evaluate the performance of the model.I used different classification models commonly used in the literature to train the proposed dataset and compared the performance using standard evaluation metrics found in the literature.

    3.2 Pornographic and Violence Level Classification Evaluation Metrics Analysis

    Fig.8 illustrates the performance comparison of the proposed model when trained using different optimizers,namely Stochastic Gradient Descent (SGD) [23] and Adam [24],as well as various learning rates(0.01,0.005,and 0.001).The objective of this comparison is to determine the optimal combination of hyperparameters for achieving the highest possible accuracy.

    When employing the SGD optimizer,the model’s performance was observed to be lower than that of the Adam optimizer.The lowest accuracy recorded with the SGD optimizer,68.6%,was obtained at a learning rate of 0.001.Conversely,the highest accuracy achieved with the SGD optimizer was 71.7%at a learning rate of 0.01.

    In contrast,using the Adam optimizer,the proposed model demonstrated marginally superior performance across all learning rates.The least accurate configuration,yielding an accuracy of 70.8%,employed the Adam optimizer at a learning rate of 0.005.Furthermore,the highest accuracy of 72.7%was attained with the Adam optimizer at a learning rate of 0.01.

    Based on these findings,the Adam optimizer with a learning rate of 0.01 was identified as the optimal combination for training the proposed model.Consequently,this configuration was selected for further evaluation and assessment of the model’s performance.

    Table 4 compares the performance of the proposed method,AnimeNet,with other state-of-the-art models,including DenseNet161 [24],ResNet50 [25],MobileNet_v3 [26],EfficientNet [27],ViT [28],ResNext50 [29],ConvNext_b [30],RegNet [31],and Wide-ResNet50 [32].The performance metrics for the methods were assessed across multiple conditions,namely EL1,EL2,EL3,NM,VL1,VL2,and VL3.

    Figure 8:The comparison of the model accuracy using different hyperparameter settings

    Table 4:The accuracy of the erotic and violence classification for different deep learning models using a 7:3 ratio

    In terms of performance across EL1,EL2,and EL3 categories,the AnimeNet model demonstrated superior performance with scores of 74.2,73.0,and 75.8,respectively,surpassing all the competing models.Notably,AnimeNet outperformed the ViT model,which scored 76.2,84.5,and 64.3 for the same categories.

    Under the NM condition,AnimeNet achieved the highest score of 84.6,followed closely by EfficientNet,with a score of 77.2.In the VL1 category,AnimeNet outperformed all other models,securing a score of 74.6.

    For VL2 and VL3 categories,AnimeNet scored 63.9 and 75.4,respectively.Despite its slightly lower performance in the VL2 category,AnimeNet still maintained a competitive stance among other models.In the VL3 category,it performed closely to DenseNet161,which achieved a score of 66.3.

    Finally,taking an overall perspective through the Accuracy metric,AnimeNet’s performance was superior to all others,with an accuracy of 74.5,affirming the robustness of the proposed method.This clearly demonstrates that our AnimeNet model stands as a significant improvement over the existing methods in multiple conditions,emphasizing its versatility and superior performance.

    To further assess the performance of the proposed model on my custom dataset,I conducted additional experiments by adjusting the training and testing data ratios to 50%each.Table 5 compares the accuracy of erotic(EL1,EL2,EL3,NM)and violence(VL1,VL2,VL3)classification for various deep learning models,including DenseNet161 [24],ResNet50 [25],MobileNet_v3 [26],EfficientNet[27],ViT[28],ResNext50[29],ConvNext_b[30],RegNet[31],Wide-ResNet50[32],and the proposed method,AnimeNet.

    Table 5:The accuracy of the erotic and violence classification for different deep learning models using a 5:5 ratio

    Among the models,AnimeNet demonstrated a superior performance across all categories.In the erotic classification categories EL1,EL2,and EL3,AnimeNet scored 70.6,75.9,and 71.9,respectively.The next closest performer in these categories was MobileNet_v3,which achieved scores of 73.0,64.5,and 64.7,respectively.In the NM category,AnimeNet again outperformed all other models,securing a score of 78.8.

    In the violence classification categories,AnimeNet consistently delivered top performance.In the VL1 category,AnimeNet achieved a score of 73.9,surpassing EfficientNet,which scored 67.6.For VL2 and VL3,AnimeNet scored 63.4 and 73.8,respectively,outperforming Wide-ResNet50,which scored 57.5 and 77.2.

    When looking at the overall accuracy(Acc),AnimeNet achieved the highest score of 72.7.The next closest was DenseNet161,with an accuracy of 68.1.The robust performance of AnimeNet across all categories affirms its superior effectiveness for erotic and violence classification,setting a new standard for similar deep-learning models.

    In order to further evaluate the performance of the proposed model on my custom dataset,I conducted additional experiments by modifying the training and testing data ratios to 30%and 70%,respectively.Table 6 presents the performance comparison of various deep learning models,including DenseNet161 [24],ResNet50 [25],MobileNet_v3 [26],EfficientNet [27],ViT [28],ResNext50 [29],ConvNext_b[30],RegNet[31],Wide-ResNet50[32],and the proposed model,AnimeNet.

    Table 6:The accuracy of the erotic and violence classification for different deep learning models using a 3:7 ratio

    In the erotic classification categories(EL1,EL2,EL3,NM),AnimeNet generally outperformed all other models.Specifically,in the EL1 category,AnimeNet achieved a performance score of 57.4,closely followed by DenseNet161,with a score of 67.1.In the EL2 category,AnimeNet stood out with the highest score of 62.6.For the EL3 and NM categories,AnimeNet achieved scores of 57.1 and 72.1,respectively,reflecting superior performance compared to the other models.

    In the violence classification categories(VL1,VL2,VL3),AnimeNet’s performance was remarkably robust.In the VL1 category,AnimeNet outperformed all other models with a score of 69.0,while in the VL2 and VL3 categories,AnimeNet achieved respectable scores of 61.8 and 71.6,respectively.Notably,in the VL3 category,AnimeNet outperformed EfficientNet,which scored 72.1.

    When considering the overall accuracy(Acc),AnimeNet outshone all other models,achieving the highest score of 65.2.The next closest model,DenseNet161,secured an accuracy of 64.8.Despite the skewed data ratio,AnimeNet’s performance remained consistent,thereby demonstrating its ability to maintain high performance in various scenarios.This underscores the efficacy of the proposed method,AnimeNet,in classifying erotic and violent content and highlights its potential applicability in real-world scenarios with uneven data distributions.

    This study also calculates the standard evaluation metrics such as precision,recall,specificity,F1-score,and accuracy for every class to evaluate the performance of the proposed model.I used a 7:3 training and testing-based model to calculate the evaluation metrics.Table 7 presents the comprehensive evaluation of the proposed erotic and violence level classification model across several standard multiclass evaluation metrics:precision,recall,specificity,F1-score,and accuracy.Each metric contributes to a holistic understanding of the model’s performance across different classes(EL1,EL2,EL3,NM,VL1,VL2,and VL3).

    Table 7:The standard multiclass evaluation metrics of the proposed erotic and violence level classification model.The precision,recall,specificity,F1-score,and accuracy are used to evaluate the model

    For class EL1,the model exhibited a precision of 74.2%,implying that when the model predicted an instance to be of class EL1,it was correct 74.2%of the time.The recall of 68.8%denotes that the model correctly identified 68.8%of all actual EL1 instances.The model’s specificity was high at 96.8%,demonstrating its ability to correctly identify instances not belonging to class EL1.The F1-score,a harmonic mean of precision and recall,was 71.4%,reflecting a balanced predictive performance.The overall accuracy for EL1 stood at 93.5%.

    The performance across other classes followed a similar pattern.For instance,the class EL2 exhibited a precision of 73.0%,a recall of 70.6%,a specificity of 96.1%,an F1-score of 71.8%,and an accuracy of 92.7%.Class EL3 demonstrated strong performance with a precision of 74.5%,recall of 76.9%,specificity of 95.9%,F1-score of 75.7%,and accuracy of 93.4%.

    For the normal (NM) class,the model had the highest performance among all classes,with a precision of 85.6%,recall of 86.1%,specificity of 97.4%,an F1-score of 85.8%,and an accuracy of 95.6%.

    For the violence categories,the model’s performance was consistently strong.Class VL1 achieved a precision of 74.6%,recall of 75.0%,specificity of 95.3%,F1-score of 74.8%,and accuracy of 92.2%.In contrast,the VL2 class showed a slight dip in performance with precision at 63.9%,recall at 60.0%,specificity at 93.8%,F1-score at 61.9%,and accuracy at 88.6%.Class VL3 rebounded with a precision of 75.4%,recall of 83.3%,specificity of 95.0%,an F1-score of 79.2%,and an accuracy of 93.2%.

    These results highlight the strong performance of the proposed model in classifying erotic and violence levels,confirming its effectiveness and reliability.

    Fig.9 presents the confusion matrix for the proposed classification model,which provides an in-depth examination of the model’s class-wise performance.This graphical representation offers a more nuanced understanding of my model’s capability and further corroborates the evaluation metrics previously discussed.

    Significantly,the confusion matrix reveals that the model achieves a comparatively higher accuracy rate for the normal(non-erotic,non-violent)class,demonstrated by fewer misclassifications.This suggests that the model has been particularly successful in recognizing and classifying images in the normal class.

    However,the confusion matrix also reveals areas where the model’s performance is less accurate.Specifically,some misclassifications occurred among the three erotic levels.This phenomenon may be attributed to the shared or similar features across these classes,which make differentiation challenging for the model.A similar trend was observed in the violence levels,with the model experiencing prediction errors between the different classes of violence.

    For example,the model incorrectly classified 40 images from violence level 2 as level 3.This type of error likely stems from the presence of overlapping or analogous patterns across these two classes.Nonetheless,such misclassifications are comparatively few,and the model demonstrates predominant success in accurately predicting the correct classes,with only a limited number of significant misclassifications.

    In summary,the overall performance of the proposed model,as demonstrated through the confusion matrix and the associated evaluation metrics,underscores its competency in undertaking the complex task of cartoon scene classification.Despite certain areas of misclassification,the model shows significant potential for practical applications within this domain.Its ability to identify and differentiate between various levels of erotic and violent content makes it a compelling choice for systems requiring precise content moderation and classification in cartoon-based media.

    Various real-world scenarios can impact the representation of cartoon images,such as zooming into scenes or cropping significant portions.Thus,assessing the robustness of a scene classification model against diverse scenarios is crucial before deploying it in real-time cartoon scene monitoring applications.In this study,I evaluate the performance of the proposed model under three common scenarios encountered in cartoon scenes,specifically cropping,rotation,and block noise,which may arise in cartoon videos.

    Fig.10 illustrates the prediction scores achieved by the proposed model when applied to randomly cropped cartoon images.The effectiveness of the model is demonstrated through a series of experiments where an original image,as shown in Fig.10a,is segmented into numerous randomly cropped subsections.These cropped images are then individually introduced into the classifier,which proceeds to predict their respective scene categories.

    Figure 10:The prediction results of the proposed model on the cropped images.(a) is the original image.(b),(c),(d),and(e)are the predicted results using the proposed model on randomly cropped violent images

    The results derived from these experiments provide strong evidence of the model’s accurate and consistent performance.Despite the challenges posed by the lack of context due to cropping,the model effectively assigns the correct scene class to each cropped image with remarkably high prediction scores.This demonstrates the model’s ability to identify salient features within each cropped section that are indicative of the overall scene category,showcasing its robustness and proficiency in managing partially obscured or fragmented images.

    Furthermore,the analysis of the prediction scores reveals that the average score across all cropped images exceeds 99%.This high rate underscores the reliability and robustness of the proposed model in processing and classifying cropped images.It suggests that even when only portions of an image are available,my model is capable of making highly accurate predictions,thereby proving its resilience to the challenges presented by partial image data.

    Further evaluations,such as rotation and block noise scenarios,are essential to comprehensively understand the model’s ability to maintain accurate classification performance despite variations in input data.A detailed analysis of these scenarios will provide additional insights into the model’s adaptability and potential for use in real-world applications.

    Fig.11 provides a visualization of the prediction scores achieved by the proposed model when applied to rotated cartoon images.This assessment unveils an intriguing relationship between the prediction accuracy of the model and the varying degree of rotation applied to the images.

    The baseline for this experiment is established with an original,unrotated image depicted in Fig.11a,which yields a prediction accuracy of 91.8%.The proposed model manages to uphold high prediction scores when images are rotated,primarily as the rotation preserves the distribution of inherent features within the image that are crucial for classification.

    Nonetheless,it is noteworthy that the model’s prediction accuracy encounters a significant dip when the images are subjected to rotations of 180 degrees and 90 degrees to the right.In these instances,the model’s prediction accuracies descend to 55.8% and 57.4%,respectively,reflecting a substantial deterioration in the model’s performance.

    Figure 11:The prediction results of the proposed model on the rotated images.(a),(b),(c),and(d)are the predicted results using the proposed model on the rotated images

    These results underline the profound influence that image rotation can have on the performance of a classification model.Despite the robustness of the model in handling a variety of alterations,rotations at certain angles significantly challenge the model’s ability to accurately predict the scene class,illustrating the model’s limitations under such conditions.

    The proposed model’s performance was further evaluated by examining its ability to classify noisy images.Fig.12 displays the prediction outcomes for images that were distorted with various block noises,demonstrating the model’s aptitude for accurate classification amidst these added challenges.

    Distinct block noises were integrated into the images in Figs.12a–12d.Despite the presence of such noise,these images were effectively processed by the proposed model,which correctly identified all of them as belonging to the violence level 1 class.Among these classifications,the lowest prediction accuracy was observed for Fig.12a,with an accuracy score of 92.7%,still a commendable level given the noisy input.

    Though the proposed model exhibits a noteworthy capacity to classify cartoon scenes even in the presence of noise,the levels of accuracy achieved in these instances were observed to be slightly lower than those attained when handling cropped images.Despite this modest reduction,the analysis of the results indicates that the proposed model maintains its robustness in the face of various forms of noise and constrained environments,characteristics that would be frequently encountered in realworld scenarios.

    Figure 12:The prediction results of the proposed model on the noisy images

    These findings,along with the prior assessments involving cropped and rotated images,contribute to a comprehensive understanding of the model’s performance under an array of diverse and challenging conditions.This collective evaluation provides a holistic view of the proposed model’s resilience and adaptability,highlighting its suitability for use in practical applications that may encompass noisy,rotated,or partially obscured imagery.

    4 Discussion

    This section focuses on evaluating various methods for detecting pornography and violence in images.Table 8 provides summarizes the detection goals,class numbers,and accuracies achieved by different authors in their respective studies.

    Table 8:Comparative analysis of AnimeNet with different existing methods for violence and pornography detection

    Peres et al.[18]addressed the goal of pornography detection and achieved an accuracy of 97.9%.Ditsanthia et al.[10]focused on violence detection,specifically targeting two classes,and obtained an accuracy of 75.73%.In contrast,Sumon et al.[11]and Ye et al.[13]also worked on violence detection with two classes but achieved higher accuracies of 97.06%and 97%,respectively.

    Yousaf et al.[14]explored the detection and classification of inappropriate cartoon content,which involved three classes.They reported an accuracy of 95.66%.Khan et al.[15] focused on violence detection similar to previous works and obtained an accuracy of 97.0%.

    Wang et al.[16]and Mohamed[17]concentrated on pornographic detection.Wang et al.achieved an accuracy of 69.24% using a two-class approach,while Mohamed achieved a higher accuracy of 94.1%with a three-class classification.

    Yuan et al.[19] contributed to violence detection with an accuracy of 95%.Shen et al.[20]specifically targeted pornographic image recognition and achieved an accuracy of 94.70%using a twoclass approach.

    In this study,I developed a method for violence and pornographic detection,considering seven different classes.My approach achieved an accuracy of 74.5%.While my method for violence and pornographic detection achieved a lower accuracy of 74.5%compared to some of the other methods discussed,it is important to provide further context and insights into my approach.

    Firstly,it is crucial to note that AnimeNet was designed to address a more complex detection task by considering seven different classes,which include different levels of violence and pornographic content.This significantly increases the difficulty of accurate classification compared to methods targeting only two or three classes.The presence of multiple classes introduces greater variability and complexity in the dataset,making it inherently more challenging to achieve high accuracies.

    Additionally,the detection of both violence and pornographic content introduces a wider range of visual characteristics and semantic cues that need to be effectively captured by the algorithm.The complexity of these detection goals requires the model to be trained on diverse and representative datasets,encompassing a wide range of potential instances and variations.

    Another factor to consider is the availability and size of the dataset used for training my model.Larger datasets with diverse and balanced samples can enhance the model’s ability to generalize and make accurate predictions.If the dataset used for training my model was relatively small or imbalanced,it could have impacted the overall performance and contributed to the lower accuracy.

    Although AnimeNet achieved a lower accuracy of 74.5%compared to some of the other methods,it is crucial to consider the complexity of the detection task,the number of classes targeted,dataset characteristics,and the challenges associated with subjectivity and context in determining inappropriate content.Despite its limitations,my method offers the advantage of simultaneously addressing violence and pornographic content detection,which can be valuable in certain applications.Future work should focus on refining the algorithm,increasing the dataset size,and exploring techniques to mitigate false positives and false negatives,ultimately aiming to improve the overall accuracy of the detection system.

    5 Conclusion

    In this paper,I present an automatic classification system designed to regulate the content of cartoons and animations by identifying violence and erotic levels.Given the unique characteristics of violent and erotic scenes,I propose AnimeNet,a novel deep-learning model capable of effectively and reliably extracting information from animated images.A novel channel-spatial attention module is introduced and integrated into the model’s block structure to extract crucial features from feature maps.

    Experimental results demonstrate that AnimeNet surpasses other deep learning models in classifying violence and erotic levels in various images.The proposed model achieved overall accuracies of 74.5%,72.7%,and 65.8% for 7:3,5:5,and 3:7 training-testing ratios,respectively.Furthermore,AnimeNet was tested in diverse noisy environments typically found in animated videos,and it maintained high performance under these challenging conditions.This showcases the model’s effectiveness in processing animated scenes efficiently.

    While the proposed method demonstrates impressive performance,there are opportunities for further improvement.This study’s limitations include the potential non-representativeness of the training dataset due to vast global animation diversity,the limited scope of the violence dataset,and the model’s untested transferability to other domains or real-world images.In the future,researchers can focus on diversifying the proposed animation dataset to enhance generalization,integrating advanced techniques to improve AnimeNet’s performance,and expanding the model’s robustness to diverse types of noise.Additionally,they may aim to test AnimeNet’s transferability to other domains and employ comprehensive evaluation metrics for a more thorough performance assessment.

    Acknowledgement:I would like to extend my sincere gratitude to Chen Yu and Sagar A S M Sharifuzzaman for their invaluable guidance and recommendations throughout the course of this research work.Their wealth of knowledge and experience,along with their continuous encouragement,has been fundamental to the successful completion of this study.

    Funding Statement:The author received no specific funding for this study.

    Author Contributions:Study conception and design: Y.Tang;data collection: Y.Tang;analysis and interpretation of results:Y.Tang;draft manuscript preparation:Y.Tang.

    Availability of Data and Materials:Data will be available upon request.

    Conflicts of Interest:The author declares that he has no conflicts of interest to report regarding the present study.

    老师上课跳d突然被开到最大视频| 高清午夜精品一区二区三区 | 欧美激情在线99| 日韩大尺度精品在线看网址| 毛片女人毛片| 日韩强制内射视频| 国产精品免费一区二区三区在线| 国产精品福利在线免费观看| 久久热精品热| 成年版毛片免费区| 一个人看视频在线观看www免费| 色尼玛亚洲综合影院| 最近的中文字幕免费完整| 好男人视频免费观看在线| 国产黄a三级三级三级人| 久久99热这里只有精品18| 欧美又色又爽又黄视频| 国产一级毛片七仙女欲春2| 亚洲成人av在线免费| 欧美不卡视频在线免费观看| 联通29元200g的流量卡| 99久久成人亚洲精品观看| 青春草国产在线视频 | 校园春色视频在线观看| 亚洲四区av| 免费搜索国产男女视频| 国产精品精品国产色婷婷| 一边亲一边摸免费视频| 久久精品国产自在天天线| 欧美bdsm另类| 搡女人真爽免费视频火全软件| 成人三级黄色视频| 91aial.com中文字幕在线观看| 中文欧美无线码| 久久热精品热| 国产亚洲5aaaaa淫片| 免费看a级黄色片| 自拍偷自拍亚洲精品老妇| 久久热精品热| 成人毛片60女人毛片免费| 精品熟女少妇av免费看| av天堂中文字幕网| 三级经典国产精品| 欧美日韩综合久久久久久| 亚洲欧美成人精品一区二区| 国产欧美日韩精品一区二区| 中文字幕久久专区| 亚洲国产精品国产精品| 欧美另类亚洲清纯唯美| 色尼玛亚洲综合影院| 成年免费大片在线观看| 国产中年淑女户外野战色| 乱系列少妇在线播放| 亚洲成人久久爱视频| 麻豆国产av国片精品| 国产片特级美女逼逼视频| 亚洲四区av| 国产成人aa在线观看| 在线免费观看不下载黄p国产| 少妇人妻精品综合一区二区 | 精品久久久久久久久久久久久| 国内久久婷婷六月综合欲色啪| 九九爱精品视频在线观看| 亚洲天堂国产精品一区在线| 99热这里只有精品一区| 欧美极品一区二区三区四区| 国产单亲对白刺激| 久久精品国产鲁丝片午夜精品| 99久久成人亚洲精品观看| 黄色欧美视频在线观看| 国产av麻豆久久久久久久| 国产精品人妻久久久久久| 亚洲无线观看免费| 伦精品一区二区三区| 色5月婷婷丁香| 不卡视频在线观看欧美| 亚洲av二区三区四区| 精品少妇黑人巨大在线播放 | 亚洲在线自拍视频| 精品久久久久久久久久免费视频| 午夜久久久久精精品| 天堂影院成人在线观看| av黄色大香蕉| 国产 一区 欧美 日韩| 97超碰精品成人国产| 亚洲性久久影院| 中文资源天堂在线| 日韩国内少妇激情av| 内射极品少妇av片p| 黄色日韩在线| 国产成人精品久久久久久| 国产三级中文精品| 亚洲国产精品sss在线观看| 久久人人精品亚洲av| 国产乱人视频| 22中文网久久字幕| 久久人人精品亚洲av| 免费人成视频x8x8入口观看| 91久久精品国产一区二区成人| av视频在线观看入口| 国产精品精品国产色婷婷| 一级毛片久久久久久久久女| 最近中文字幕高清免费大全6| 久久韩国三级中文字幕| 亚洲aⅴ乱码一区二区在线播放| 校园人妻丝袜中文字幕| а√天堂www在线а√下载| 免费观看的影片在线观看| 国产精品一区www在线观看| 人妻少妇偷人精品九色| 午夜激情福利司机影院| 国产精品一二三区在线看| 女的被弄到高潮叫床怎么办| 中出人妻视频一区二区| 午夜福利成人在线免费观看| 久久人妻av系列| 麻豆一二三区av精品| 一个人看视频在线观看www免费| 免费av不卡在线播放| 看十八女毛片水多多多| 在线天堂最新版资源| 欧美日韩一区二区视频在线观看视频在线 | 一级二级三级毛片免费看| 免费看美女性在线毛片视频| 亚洲欧美日韩卡通动漫| 欧美成人一区二区免费高清观看| 午夜免费激情av| 亚洲欧美中文字幕日韩二区| 精品人妻熟女av久视频| 亚洲精品456在线播放app| 国产高清三级在线| a级一级毛片免费在线观看| 亚洲最大成人中文| 白带黄色成豆腐渣| 国产极品精品免费视频能看的| 男女啪啪激烈高潮av片| 婷婷色综合大香蕉| 男女做爰动态图高潮gif福利片| 色综合亚洲欧美另类图片| 国产一区亚洲一区在线观看| a级毛片a级免费在线| 性插视频无遮挡在线免费观看| 欧美色视频一区免费| 久久人人爽人人片av| 久久草成人影院| av在线老鸭窝| a级一级毛片免费在线观看| 国产午夜精品一二区理论片| 伊人久久精品亚洲午夜| 你懂的网址亚洲精品在线观看 | 不卡视频在线观看欧美| 成人亚洲欧美一区二区av| 亚洲av中文av极速乱| 久久国产乱子免费精品| 久久精品久久久久久噜噜老黄 | 91av网一区二区| 日韩一本色道免费dvd| 日韩在线高清观看一区二区三区| 黄色一级大片看看| 人妻夜夜爽99麻豆av| 亚洲精品自拍成人| 国产精品福利在线免费观看| 久久久久久久久大av| 有码 亚洲区| 天美传媒精品一区二区| 性色avwww在线观看| 久久99热6这里只有精品| 国产av一区在线观看免费| 国产激情偷乱视频一区二区| 人人妻人人澡人人爽人人夜夜 | 哪里可以看免费的av片| 欧美日韩乱码在线| 亚洲精品久久国产高清桃花| 亚洲av免费高清在线观看| 婷婷色av中文字幕| 久久精品国产鲁丝片午夜精品| a级一级毛片免费在线观看| 国产精品一区二区三区四区免费观看| 国产黄片美女视频| 又粗又爽又猛毛片免费看| 免费不卡的大黄色大毛片视频在线观看 | 久久久色成人| 精品久久久久久成人av| 中文字幕久久专区| 亚洲欧洲国产日韩| 国产高清视频在线观看网站| 日本av手机在线免费观看| 插阴视频在线观看视频| 99热全是精品| 国产高清三级在线| 国产精品乱码一区二三区的特点| 熟女电影av网| 日本与韩国留学比较| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 搞女人的毛片| 91久久精品电影网| 91aial.com中文字幕在线观看| 在线免费观看的www视频| 青春草亚洲视频在线观看| 在线观看午夜福利视频| 久久精品国产亚洲网站| 精品久久久噜噜| 国产精品国产高清国产av| 久久欧美精品欧美久久欧美| 国产精品人妻久久久影院| 青青草视频在线视频观看| 成人无遮挡网站| 91精品国产九色| 中文在线观看免费www的网站| 国产亚洲精品av在线| 在线观看66精品国产| 欧美日本亚洲视频在线播放| av又黄又爽大尺度在线免费看 | 日本色播在线视频| 亚洲无线在线观看| 久久久久网色| 天堂影院成人在线观看| 亚洲色图av天堂| 久久人人精品亚洲av| 99在线人妻在线中文字幕| 亚洲精品乱码久久久v下载方式| 桃色一区二区三区在线观看| 少妇熟女aⅴ在线视频| 亚洲18禁久久av| 国内精品宾馆在线| 色综合色国产| 18禁在线播放成人免费| 午夜精品一区二区三区免费看| 久久99蜜桃精品久久| 国产成人影院久久av| 国产亚洲精品av在线| 亚洲av.av天堂| 青春草国产在线视频 | av国产免费在线观看| 99久久精品国产国产毛片| 欧美日本视频| 有码 亚洲区| 免费黄网站久久成人精品| 大型黄色视频在线免费观看| 男人的好看免费观看在线视频| 亚洲无线观看免费| 国产一区二区激情短视频| 欧美高清性xxxxhd video| 男女啪啪激烈高潮av片| 亚洲最大成人手机在线| 欧美不卡视频在线免费观看| 黄色视频,在线免费观看| 国产色爽女视频免费观看| 深夜a级毛片| 久久九九热精品免费| 日本熟妇午夜| av在线天堂中文字幕| 日本一本二区三区精品| 26uuu在线亚洲综合色| 午夜爱爱视频在线播放| 成人特级黄色片久久久久久久| 一个人免费在线观看电影| 中文字幕久久专区| 色吧在线观看| 欧美高清性xxxxhd video| 久久欧美精品欧美久久欧美| 国产精品久久久久久精品电影小说 | 国产精品野战在线观看| 深夜精品福利| 中文字幕人妻熟人妻熟丝袜美| 自拍偷自拍亚洲精品老妇| 日本欧美国产在线视频| 人体艺术视频欧美日本| 一个人观看的视频www高清免费观看| 看黄色毛片网站| 色综合亚洲欧美另类图片| 国产亚洲欧美98| 身体一侧抽搐| 国产精华一区二区三区| 夫妻性生交免费视频一级片| 亚洲aⅴ乱码一区二区在线播放| 在线国产一区二区在线| 国产亚洲91精品色在线| 国产色爽女视频免费观看| 国产不卡一卡二| 99热全是精品| 男女啪啪激烈高潮av片| 一本一本综合久久| 日本成人三级电影网站| av卡一久久| 欧美日韩国产亚洲二区| 国产一区二区三区在线臀色熟女| 国产伦精品一区二区三区四那| av免费在线看不卡| 国产av一区在线观看免费| 少妇人妻一区二区三区视频| 国产美女午夜福利| 成人毛片60女人毛片免费| 嫩草影院入口| h日本视频在线播放| 青青草视频在线视频观看| 成人欧美大片| 波多野结衣高清作品| 日本在线视频免费播放| 久久精品夜夜夜夜夜久久蜜豆| 国产日本99.免费观看| 国产午夜福利久久久久久| 欧美成人a在线观看| 国产免费男女视频| 偷拍熟女少妇极品色| 日韩一区二区视频免费看| 99久久中文字幕三级久久日本| 精品日产1卡2卡| 亚洲真实伦在线观看| 午夜爱爱视频在线播放| 99久久久亚洲精品蜜臀av| 18禁裸乳无遮挡免费网站照片| 人妻制服诱惑在线中文字幕| 偷拍熟女少妇极品色| 看片在线看免费视频| 国产成年人精品一区二区| 久久精品国产亚洲av天美| 亚洲欧美日韩无卡精品| 女人被狂操c到高潮| 久久欧美精品欧美久久欧美| 最近的中文字幕免费完整| 日本黄大片高清| 最后的刺客免费高清国语| 国产精品一区二区三区四区免费观看| 国产高清有码在线观看视频| 亚洲最大成人中文| 亚洲七黄色美女视频| 日日摸夜夜添夜夜添av毛片| 噜噜噜噜噜久久久久久91| 可以在线观看毛片的网站| 在线免费观看的www视频| 日本在线视频免费播放| 菩萨蛮人人尽说江南好唐韦庄 | 级片在线观看| av免费在线看不卡| 久久人妻av系列| 99久久精品热视频| 高清在线视频一区二区三区 | 亚洲av免费在线观看| 高清在线视频一区二区三区 | 波多野结衣高清作品| 乱码一卡2卡4卡精品| 草草在线视频免费看| 中文字幕制服av| 日韩欧美精品v在线| 麻豆精品久久久久久蜜桃| 精品不卡国产一区二区三区| 夜夜爽天天搞| 在线观看66精品国产| 久久韩国三级中文字幕| 婷婷精品国产亚洲av| 丝袜美腿在线中文| 国语自产精品视频在线第100页| 最近中文字幕高清免费大全6| 亚洲无线观看免费| 天美传媒精品一区二区| 看免费成人av毛片| 国产高清有码在线观看视频| 少妇高潮的动态图| 18禁在线无遮挡免费观看视频| 国产精品一区二区三区四区免费观看| 一个人观看的视频www高清免费观看| 九草在线视频观看| 国产精品久久久久久av不卡| 欧美日韩综合久久久久久| 午夜福利在线观看吧| 欧美最新免费一区二区三区| 丝袜美腿在线中文| 看片在线看免费视频| 男女啪啪激烈高潮av片| 高清午夜精品一区二区三区 | 亚洲一区高清亚洲精品| 日本-黄色视频高清免费观看| a级毛片免费高清观看在线播放| 亚洲欧美精品综合久久99| 亚洲精品乱码久久久久久按摩| 在现免费观看毛片| 小说图片视频综合网站| av在线播放精品| av在线亚洲专区| 欧美性猛交黑人性爽| 国产探花在线观看一区二区| 国产精品免费一区二区三区在线| 亚洲不卡免费看| 亚洲精品成人久久久久久| 欧美色欧美亚洲另类二区| 12—13女人毛片做爰片一| 毛片女人毛片| 欧美日韩综合久久久久久| 免费人成视频x8x8入口观看| 国产探花极品一区二区| 九九热线精品视视频播放| 国产探花在线观看一区二区| 亚洲av中文av极速乱| 欧美zozozo另类| 国产探花在线观看一区二区| 蜜臀久久99精品久久宅男| 六月丁香七月| 啦啦啦韩国在线观看视频| 欧美日本亚洲视频在线播放| 久久九九热精品免费| 少妇熟女aⅴ在线视频| 久久精品综合一区二区三区| 直男gayav资源| 亚洲真实伦在线观看| 精品欧美国产一区二区三| 国产精品乱码一区二三区的特点| 欧美精品国产亚洲| 久久精品国产亚洲av天美| 国产极品天堂在线| 毛片女人毛片| 久久久久久久久久久免费av| 青春草亚洲视频在线观看| 亚洲欧美日韩高清在线视频| av国产免费在线观看| 校园春色视频在线观看| av免费在线看不卡| 一级毛片aaaaaa免费看小| 九九热线精品视视频播放| 亚洲无线观看免费| 99久久无色码亚洲精品果冻| 高清午夜精品一区二区三区 | 最好的美女福利视频网| 晚上一个人看的免费电影| 精品国产三级普通话版| 又黄又爽又刺激的免费视频.| 亚洲av成人av| 亚洲国产日韩欧美精品在线观看| 91在线精品国自产拍蜜月| h日本视频在线播放| 一本久久精品| 日韩人妻高清精品专区| 久久草成人影院| 亚洲美女视频黄频| 亚洲欧美日韩高清在线视频| 嫩草影院精品99| 一个人看视频在线观看www免费| 欧美zozozo另类| 精品不卡国产一区二区三区| 成人性生交大片免费视频hd| 亚洲成人精品中文字幕电影| 国产淫片久久久久久久久| 日本成人三级电影网站| 亚洲国产色片| 日本与韩国留学比较| 中文字幕制服av| 亚洲综合色惰| 日韩一区二区视频免费看| 一级二级三级毛片免费看| 欧美激情久久久久久爽电影| 春色校园在线视频观看| 日韩人妻高清精品专区| 99久国产av精品| 亚洲av男天堂| 天堂av国产一区二区熟女人妻| 欧美性感艳星| 欧美最新免费一区二区三区| 国产精品久久久久久精品电影小说 | 边亲边吃奶的免费视频| 99热这里只有是精品在线观看| 3wmmmm亚洲av在线观看| 久久久a久久爽久久v久久| 亚洲精品粉嫩美女一区| 国产伦理片在线播放av一区 | 18+在线观看网站| 亚洲七黄色美女视频| 人体艺术视频欧美日本| 国产麻豆成人av免费视频| 亚洲自拍偷在线| 直男gayav资源| 长腿黑丝高跟| 国产免费男女视频| 能在线免费看毛片的网站| 边亲边吃奶的免费视频| 亚洲精品456在线播放app| 国产成人91sexporn| 黄色欧美视频在线观看| 青青草视频在线视频观看| 亚洲经典国产精华液单| 国产老妇伦熟女老妇高清| 久久99精品国语久久久| 国内久久婷婷六月综合欲色啪| 长腿黑丝高跟| 国产探花极品一区二区| 男插女下体视频免费在线播放| 中文字幕av成人在线电影| 91精品国产九色| 色吧在线观看| 亚洲精华国产精华液的使用体验 | 午夜福利在线观看免费完整高清在 | 在线观看免费视频日本深夜| 国产成人aa在线观看| av国产免费在线观看| 丝袜喷水一区| 18禁在线播放成人免费| 九九在线视频观看精品| 久久精品国产亚洲网站| www.色视频.com| 午夜精品一区二区三区免费看| 中文字幕熟女人妻在线| 三级男女做爰猛烈吃奶摸视频| 人妻少妇偷人精品九色| 黄片wwwwww| 天堂网av新在线| 欧美性感艳星| 中文亚洲av片在线观看爽| 热99re8久久精品国产| 久久精品久久久久久噜噜老黄 | 欧美日韩一区二区视频在线观看视频在线 | 国产在线精品亚洲第一网站| 丰满乱子伦码专区| 99久久九九国产精品国产免费| 免费观看的影片在线观看| 男女啪啪激烈高潮av片| 亚洲久久久久久中文字幕| 哪里可以看免费的av片| 少妇熟女aⅴ在线视频| 日本一二三区视频观看| 亚洲欧美精品专区久久| 乱码一卡2卡4卡精品| 亚洲最大成人av| 欧美最新免费一区二区三区| 午夜激情欧美在线| 日韩制服骚丝袜av| 久久久久久久久中文| 性插视频无遮挡在线免费观看| a级毛片免费高清观看在线播放| 特级一级黄色大片| 国产精华一区二区三区| 欧美日韩在线观看h| 国产高清三级在线| 一个人免费在线观看电影| 亚洲国产高清在线一区二区三| 国产视频内射| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 国产一区二区在线观看日韩| 国产探花在线观看一区二区| 久久6这里有精品| 亚洲成人久久性| 中文字幕制服av| 床上黄色一级片| 老熟妇乱子伦视频在线观看| 国产成人aa在线观看| 精品国产三级普通话版| 波多野结衣巨乳人妻| 性插视频无遮挡在线免费观看| 久久久久久九九精品二区国产| 精品久久久久久成人av| 亚洲内射少妇av| 欧美不卡视频在线免费观看| 最近2019中文字幕mv第一页| 男女下面进入的视频免费午夜| 一级毛片aaaaaa免费看小| 日韩av在线大香蕉| 国产爱豆传媒在线观看| 99久久人妻综合| 久久久精品欧美日韩精品| 变态另类成人亚洲欧美熟女| 麻豆乱淫一区二区| 国产精品国产三级国产av玫瑰| 欧美人与善性xxx| 久久午夜亚洲精品久久| av免费观看日本| 精品国内亚洲2022精品成人| 免费人成视频x8x8入口观看| 亚洲精品亚洲一区二区| 精品久久久久久久久久久久久| 黄色视频,在线免费观看| 亚洲国产高清在线一区二区三| 又爽又黄无遮挡网站| 天堂√8在线中文| 亚洲色图av天堂| 秋霞在线观看毛片| 一本—道久久a久久精品蜜桃钙片 精品乱码久久久久久99久播 | 久久久成人免费电影| 国产激情偷乱视频一区二区| 久久午夜福利片| av在线观看视频网站免费| 国产精品麻豆人妻色哟哟久久 | 欧美一区二区国产精品久久精品| 夫妻性生交免费视频一级片| 久久99热6这里只有精品| 国产成年人精品一区二区| 成人高潮视频无遮挡免费网站| 可以在线观看的亚洲视频| www.色视频.com| 久久久久久久久久成人| 国产老妇伦熟女老妇高清| 亚洲欧洲国产日韩| 欧美日本视频| 好男人在线观看高清免费视频| 3wmmmm亚洲av在线观看| 赤兔流量卡办理| 男插女下体视频免费在线播放| 精华霜和精华液先用哪个| 精品久久久久久久人妻蜜臀av| 99riav亚洲国产免费| 哪个播放器可以免费观看大片| 日韩一区二区视频免费看| 一边摸一边抽搐一进一小说| 国产真实伦视频高清在线观看| 一级毛片aaaaaa免费看小| 中国美女看黄片| 亚洲国产精品sss在线观看| 日韩一区二区视频免费看| 一级毛片我不卡| av国产免费在线观看| 久久久久久久久中文| 给我免费播放毛片高清在线观看| 少妇丰满av| 亚洲国产日韩欧美精品在线观看| 日本欧美国产在线视频| 亚洲美女视频黄频| 国产伦一二天堂av在线观看| 简卡轻食公司|