Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00059
Chutian Jiang
As model prediction becomes more and more accurate and the network becomes deeper and deeper, the amount of memory consumed by the neural network becomes a problem, especially on mobile devices. It is also very difficult to balance the tradeoff between computational cost and battery life, which makes mobile devices very hard as well to become smarter. Model quantification techniques provide the opportunity to tackle this tradeoff by reducing the memory bandwidth and storage and improving the system throughput and latency. This paper discusses and compares the state-of-the-art methods of neural network quantification methodologies including Post Training Quantization (PTQ) and Quantization Aware Training (QAT). PTQ directly quantizes the trained floating-point model. The implementation process is simple and does not require quantization during the training phase. QAT requires us to use simulated quantization operations to model the effect of the quantization, and forward and backward passes are usually performed in the floating-point model. Finally, as discussed in the experiments in this paper, we conclude that with the evolution of the quantization techniques, the accuracy gap between PTQ and QAT is shrinking.
{"title":"Efficient Quantization Techniques for Deep Neural Networks","authors":"Chutian Jiang","doi":"10.1109/CONF-SPML54095.2021.00059","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00059","url":null,"abstract":"As model prediction becomes more and more accurate and the network becomes deeper and deeper, the amount of memory consumed by the neural network becomes a problem, especially on mobile devices. It is also very difficult to balance the tradeoff between computational cost and battery life, which makes mobile devices very hard as well to become smarter. Model quantification techniques provide the opportunity to tackle this tradeoff by reducing the memory bandwidth and storage and improving the system throughput and latency. This paper discusses and compares the state-of-the-art methods of neural network quantification methodologies including Post Training Quantization (PTQ) and Quantization Aware Training (QAT). PTQ directly quantizes the trained floating-point model. The implementation process is simple and does not require quantization during the training phase. QAT requires us to use simulated quantization operations to model the effect of the quantization, and forward and backward passes are usually performed in the floating-point model. Finally, as discussed in the experiments in this paper, we conclude that with the evolution of the quantization techniques, the accuracy gap between PTQ and QAT is shrinking.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"33 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134005528","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00041
Hao Bai
Action recognition and pose estimation from videos are closely related to understand human motions, but more literature focuses on how to solve pose estimation tasks alone from action recognition. This research shows a faster and more flexible training method for VideoPose3D which is based on action recognition. This model is fed with the same type of action as the type that will be estimated, and different types of actions can be trained separately. Evidence has shown that, for common pose-estimation tasks, this model requires a relatively small amount of data to carry out similar results with the original research, and for action-oriented tasks, it outperforms the original research by 4.5% with a limited receptive field size and training epoch on Velocity Error of MPJPE. This model can handle both action-oriented and common pose-estimation problems.
{"title":"A Training Method For VideoPose3D with Ideology of Action Recognition","authors":"Hao Bai","doi":"10.1109/CONF-SPML54095.2021.00041","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00041","url":null,"abstract":"Action recognition and pose estimation from videos are closely related to understand human motions, but more literature focuses on how to solve pose estimation tasks alone from action recognition. This research shows a faster and more flexible training method for VideoPose3D which is based on action recognition. This model is fed with the same type of action as the type that will be estimated, and different types of actions can be trained separately. Evidence has shown that, for common pose-estimation tasks, this model requires a relatively small amount of data to carry out similar results with the original research, and for action-oriented tasks, it outperforms the original research by 4.5% with a limited receptive field size and training epoch on Velocity Error of MPJPE. This model can handle both action-oriented and common pose-estimation problems.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121858853","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00043
Danyating Shen, Takara E. Truong, C. Weintz
The prevalence of COVID-19 has illuminated the need for practical digital education tools over the past year. With students studying from home, teachers have struggled to provide their students with adequately challenging coursework. Our project aims to solve this issue in the context of math. More specifically, our goal is to encourage thoughtful learning by supplying students with personalized two-number addition problems that take time to solve but expect the student to answer correctly still. Our solution is to model the process of selecting a math problem to give a student as a Markov Decision Process (MDP) and then use Q-learning to determine the best policy for arriving at the most optimally challenging two-number addition problem for that student. The project creates three student simulators based on group member data. We show that it took student one: $(162 pm 134)$ iterations to give appropriate level problems where the first entry is mean and the second is the standard deviation. Student two took $(230 pm 205)$ iterations, and student three took $(247 pm 236)$ iterations. Lastly, we demonstrate that pre-training our model on students two and three and testing on student one showed a significant improvement from $(162 pm 134)$ iterations to $(35 pm 44)$ iterations.
{"title":"Using Q-Learning to Personalize Pedagogical Policies for Addition Problems","authors":"Danyating Shen, Takara E. Truong, C. Weintz","doi":"10.1109/CONF-SPML54095.2021.00043","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00043","url":null,"abstract":"The prevalence of COVID-19 has illuminated the need for practical digital education tools over the past year. With students studying from home, teachers have struggled to provide their students with adequately challenging coursework. Our project aims to solve this issue in the context of math. More specifically, our goal is to encourage thoughtful learning by supplying students with personalized two-number addition problems that take time to solve but expect the student to answer correctly still. Our solution is to model the process of selecting a math problem to give a student as a Markov Decision Process (MDP) and then use Q-learning to determine the best policy for arriving at the most optimally challenging two-number addition problem for that student. The project creates three student simulators based on group member data. We show that it took student one: $(162 pm 134)$ iterations to give appropriate level problems where the first entry is mean and the second is the standard deviation. Student two took $(230 pm 205)$ iterations, and student three took $(247 pm 236)$ iterations. Lastly, we demonstrate that pre-training our model on students two and three and testing on student one showed a significant improvement from $(162 pm 134)$ iterations to $(35 pm 44)$ iterations.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"2010 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123923812","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00031
Junyi Cao
Speech emotion recognition has been widely used in recent years and has become a heated topic for research. Focused on the convolutional neural network model using spectrograms as input, the CNN-LSTM model based on feature vectors, original speech signal and Log-mel spectrograms, the performance of different models is compared as well as analyzed. The study found that there are some common problems existing in the classification performance of the model. The features and algorithms currently used can effectively distinguish emotions with varied “arousal”, but it is difficult to identify the feelings with similar arousal, among the models. The CNN-LSTM model with Log-mel spectrograms as input achieved the highest accuracy.
{"title":"Application of Machine Learning Algorithms in Speech Emotion Recognition","authors":"Junyi Cao","doi":"10.1109/CONF-SPML54095.2021.00031","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00031","url":null,"abstract":"Speech emotion recognition has been widely used in recent years and has become a heated topic for research. Focused on the convolutional neural network model using spectrograms as input, the CNN-LSTM model based on feature vectors, original speech signal and Log-mel spectrograms, the performance of different models is compared as well as analyzed. The study found that there are some common problems existing in the classification performance of the model. The features and algorithms currently used can effectively distinguish emotions with varied “arousal”, but it is difficult to identify the feelings with similar arousal, among the models. The CNN-LSTM model with Log-mel spectrograms as input achieved the highest accuracy.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"388 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132495963","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00072
Yixin Liu
This paper aims to get good trajectory tracking performance for 1-DOP(degrees of freedom) manipulator system driven by pneumatic muscles. However, it is difficult for achieving wonderful trajectory tracking performance due to nonlinearity of the 1-DOP manipulator system. The integral sliding mode trajectory tracking method is shown on 1-DOP manipulator system within the paper. A nonlinear extended state observer is proposed for estimating the nonlinearity of 1-DOP manipulator system. Moreover, an integral sliding mode controller on the basis of nonlinear extended state observer is adopted for getting good trajectory tracking performance in 1-DOP manipulator system. Finally, results of the simulation show that good trajectory tracking performance is achieved by proposed integral sliding mode trajectory tracking method within the paper.
{"title":"Integral Sliding Mode Trajectory Tracking Method for 1-DOP Manipulator Systems driven by Pneumatic Muscles on the Basis of the Nonlinear Extended State Observer","authors":"Yixin Liu","doi":"10.1109/CONF-SPML54095.2021.00072","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00072","url":null,"abstract":"This paper aims to get good trajectory tracking performance for 1-DOP(degrees of freedom) manipulator system driven by pneumatic muscles. However, it is difficult for achieving wonderful trajectory tracking performance due to nonlinearity of the 1-DOP manipulator system. The integral sliding mode trajectory tracking method is shown on 1-DOP manipulator system within the paper. A nonlinear extended state observer is proposed for estimating the nonlinearity of 1-DOP manipulator system. Moreover, an integral sliding mode controller on the basis of nonlinear extended state observer is adopted for getting good trajectory tracking performance in 1-DOP manipulator system. Finally, results of the simulation show that good trajectory tracking performance is achieved by proposed integral sliding mode trajectory tracking method within the paper.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116868265","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00015
Jiahao Liang
Recommender Systems have been generally utilized in different areas including motion pictures, news, music with an intend to give the most important recommendations to clients from an assortment of accessible alternatives. Recommender Systems are planned utilizing procedures from numerous fields, some of which are: AI, data recovery, information mining, direct variable based math and man-made consciousness. However, in typical commodity applications, due to the huge user and project library and just few evaluations (Sparsity issue), and at the point when the client is new to Recommender Frameworks, the framework can’t prescribe things that are applicable to clients in light of absence of past data about the client as well as the client thing rating history that assists with deciding the clients’ preferences (cold start). What’s more, presently there’s an innovation called Context-aware Recommender Systems (CARS), which utilizing setting information (location, time, peer, etc.) during the time spent proposal. In this work, we present an outline of some of noticeable customary RS and the high level CARS. We discuss the advantages and disadvantages of them. Furthermore, we reveal some inherent problems in RS. At last, we make a conclusion and give some challenges in current works.
{"title":"An Overview of Recommender Systems and Its Next Generation: Context-Aware Recommender Systems","authors":"Jiahao Liang","doi":"10.1109/CONF-SPML54095.2021.00015","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00015","url":null,"abstract":"Recommender Systems have been generally utilized in different areas including motion pictures, news, music with an intend to give the most important recommendations to clients from an assortment of accessible alternatives. Recommender Systems are planned utilizing procedures from numerous fields, some of which are: AI, data recovery, information mining, direct variable based math and man-made consciousness. However, in typical commodity applications, due to the huge user and project library and just few evaluations (Sparsity issue), and at the point when the client is new to Recommender Frameworks, the framework can’t prescribe things that are applicable to clients in light of absence of past data about the client as well as the client thing rating history that assists with deciding the clients’ preferences (cold start). What’s more, presently there’s an innovation called Context-aware Recommender Systems (CARS), which utilizing setting information (location, time, peer, etc.) during the time spent proposal. In this work, we present an outline of some of noticeable customary RS and the high level CARS. We discuss the advantages and disadvantages of them. Furthermore, we reveal some inherent problems in RS. At last, we make a conclusion and give some challenges in current works.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"47 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115674781","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00022
Xuanang Chen
Mechatronics system design is more than just integrating electronic, software, and mechanical design, the additional features must be contingent on the ability of the mechatronic designer to optimize a design solution through these disparate fields. The basic idea of Mechatronics modeling is to divide the system into several subsystems according to the mathematical or physical characteristics of the system and the functions to be realized, and then use a variety of domain means to model according to the various links involved in these subsystems. This paper will introduce a common linear mechatronics modeling method based on a variety of simulation software.
{"title":"Mechatronics: System Analysis Based on Software Simulation and Programming","authors":"Xuanang Chen","doi":"10.1109/CONF-SPML54095.2021.00022","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00022","url":null,"abstract":"Mechatronics system design is more than just integrating electronic, software, and mechanical design, the additional features must be contingent on the ability of the mechatronic designer to optimize a design solution through these disparate fields. The basic idea of Mechatronics modeling is to divide the system into several subsystems according to the mathematical or physical characteristics of the system and the functions to be realized, and then use a variety of domain means to model according to the various links involved in these subsystems. This paper will introduce a common linear mechatronics modeling method based on a variety of simulation software.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"74 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126083166","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00023
Z. Luo
Noises are inevitable in images and videos. Many denoising algorithms have been proposed. As the acquirement of image and video qualities gets higher, algorithms with high performances and easy implementation are the new trend. With the development of deep learning, neural network has been applied in denoising algorithms. These methods represent better performances and obtain video of high quality. In this paper, we will discuss several denoising methods for both images and videos on their architectures, analyze their features and comment. Finally, we will carefully forecast what neural network based denoising modules can be built in the future study.
{"title":"Research on Neural Network Based Image and Video Denoising","authors":"Z. Luo","doi":"10.1109/CONF-SPML54095.2021.00023","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00023","url":null,"abstract":"Noises are inevitable in images and videos. Many denoising algorithms have been proposed. As the acquirement of image and video qualities gets higher, algorithms with high performances and easy implementation are the new trend. With the development of deep learning, neural network has been applied in denoising algorithms. These methods represent better performances and obtain video of high quality. In this paper, we will discuss several denoising methods for both images and videos on their architectures, analyze their features and comment. Finally, we will carefully forecast what neural network based denoising modules can be built in the future study.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128291966","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00039
Ziyuan Ma
With the continuous development of computer technology, the amount of data shared on the Internet has increased significantly. Increasing demand for big data continues to grow, facing many security challenges. In today’s information era, data on the Internet is vulnerable to various attacks, and everyone wants to protect their privacy. Therefore, maintaining the security of user-class data has become the current research hotspot. This article integrates four methods, the traditional two-prime RSA, two-prime mixed CRT, and three-prime CRT-RSA and traditional tetraprime and optimized quatertraprime methods operate on the same message summaries to record their respective time-consuming behaviors. optimization is of great significance.
{"title":"Comparative Study of the Optimization of the Multi-prime RSA Algorithm","authors":"Ziyuan Ma","doi":"10.1109/CONF-SPML54095.2021.00039","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00039","url":null,"abstract":"With the continuous development of computer technology, the amount of data shared on the Internet has increased significantly. Increasing demand for big data continues to grow, facing many security challenges. In today’s information era, data on the Internet is vulnerable to various attacks, and everyone wants to protect their privacy. Therefore, maintaining the security of user-class data has become the current research hotspot. This article integrates four methods, the traditional two-prime RSA, two-prime mixed CRT, and three-prime CRT-RSA and traditional tetraprime and optimized quatertraprime methods operate on the same message summaries to record their respective time-consuming behaviors. optimization is of great significance.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128493264","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2021-11-01DOI: 10.1109/CONF-SPML54095.2021.00070
Fan Zhang
Zero-Shot Learning aims to recognize unseen class images that do not appear in training, which is attracting more and more research interests in recently years. Side information is an important key to ZSL since it transfers the knowledge between seen and unseen classes. Human annotated attribute, as the most popular side information, need much human effort and time consumption during data collection. While unsupervised side information such as word2vec is not performing well since they lack the representation ability for visual information. In this paper, we propose to use CLIP features, which is learned with image and natural language pairs without human efforts, to perform ZSL. Extensive experiments on two benchmark datasets, AWA2 and CUB, demonstrates that our method is achieving impressive accuracy gain over word2vec, even beats human attributes in some circumstances.
{"title":"Learning Unsupervised Side Information for Zero-Shot Learning","authors":"Fan Zhang","doi":"10.1109/CONF-SPML54095.2021.00070","DOIUrl":"https://doi.org/10.1109/CONF-SPML54095.2021.00070","url":null,"abstract":"Zero-Shot Learning aims to recognize unseen class images that do not appear in training, which is attracting more and more research interests in recently years. Side information is an important key to ZSL since it transfers the knowledge between seen and unseen classes. Human annotated attribute, as the most popular side information, need much human effort and time consumption during data collection. While unsupervised side information such as word2vec is not performing well since they lack the representation ability for visual information. In this paper, we propose to use CLIP features, which is learned with image and natural language pairs without human efforts, to perform ZSL. Extensive experiments on two benchmark datasets, AWA2 and CUB, demonstrates that our method is achieving impressive accuracy gain over word2vec, even beats human attributes in some circumstances.","PeriodicalId":415094,"journal":{"name":"2021 International Conference on Signal Processing and Machine Learning (CONF-SPML)","volume":"33 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124596953","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：