Pub Date : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543389
Yuantian Zhu
The rapid development of science and technology in our country has led to the rapid development of Chinese enterprises, and the scale of enterprise production has continued to expand. However, as enterprises continue to accelerate their expansion and development, it follows that a huge scale is generated during the operation of the enterprise. The amount of data is increasing, and the hidden dangers of enterprises are also escalating. With the continuous competition among enterprises, the predictive and early warning technology under artificial intelligence has become the most advantageous competitiveness among enterprises. At the same time, the immeasurable losses caused by risks have made enterprises' desire for artificial intelligence monitoring and early warning technology more intense and urgent. Nowadays, most companies are still using more traditional corporate management methods. This traditional management method has many problems: mostly based on experience and visual inspection, thus ignoring the attention to some risks that cannot be visually observed, and cannot pay attention to the current corporate risks. Quantitative analysis and evaluation of the status can not achieve the effect of accurately preventing risks, so that the potential value of a large amount of data cannot be fully explored.
{"title":"Research on Dynamic Monitoring and Early Warning Methods of Company Management Driven by Artificial Intelligence","authors":"Yuantian Zhu","doi":"10.1109/CSAIEE54046.2021.9543389","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543389","url":null,"abstract":"The rapid development of science and technology in our country has led to the rapid development of Chinese enterprises, and the scale of enterprise production has continued to expand. However, as enterprises continue to accelerate their expansion and development, it follows that a huge scale is generated during the operation of the enterprise. The amount of data is increasing, and the hidden dangers of enterprises are also escalating. With the continuous competition among enterprises, the predictive and early warning technology under artificial intelligence has become the most advantageous competitiveness among enterprises. At the same time, the immeasurable losses caused by risks have made enterprises' desire for artificial intelligence monitoring and early warning technology more intense and urgent. Nowadays, most companies are still using more traditional corporate management methods. This traditional management method has many problems: mostly based on experience and visual inspection, thus ignoring the attention to some risks that cannot be visually observed, and cannot pay attention to the current corporate risks. Quantitative analysis and evaluation of the status can not achieve the effect of accurately preventing risks, so that the potential value of a large amount of data cannot be fully explored.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"37 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"117139275","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543161
C. Chen
People nowadays search for answering the questions on Q&A platforms online such as Zhihu and Quora. As many rely on these platforms, filtering controversial questions, including but not limited to hate speeches and online racism, is particularly important. While human resources are too scarce, using Artificial Intelligence to filter out some disputable and insulting questions is essential. In this work, we propose a deep learning-based classification method to analyze the sincerity of questions from Quora and achieve an overall 95.25% accuracy.
{"title":"Insincere Question Classification by Deep Neural Networks","authors":"C. Chen","doi":"10.1109/CSAIEE54046.2021.9543161","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543161","url":null,"abstract":"People nowadays search for answering the questions on Q&A platforms online such as Zhihu and Quora. As many rely on these platforms, filtering controversial questions, including but not limited to hate speeches and online racism, is particularly important. While human resources are too scarce, using Artificial Intelligence to filter out some disputable and insulting questions is essential. In this work, we propose a deep learning-based classification method to analyze the sincerity of questions from Quora and achieve an overall 95.25% accuracy.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"93 10 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125018093","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543310
Xiao-ling Ren, Deyi Yang
With the continuous development of the times, intelligent teaching assistance is also attracting more and more attention in education. The intelligent detection algorithm for student behavior is gradually becoming more precise. In university classrooms, students sleep on mobile phones and are more serious. Intelligent education can more accurately identify student behaviors to help teachers optimize teaching methods, thereby improving students' classroom learning effects. This paper studies and improves YOLOv4, and proposes a network structure called YOLOv4-Bi, which mainly adds the enhanced feature extraction network of YOLOv4 to the feature extraction structure of jumping and top-down, bottom-up combined paths. The used student classroom recording video is enhanced by taking the frame data and training, and testing on this data set. The original YOLOv4 is compared with the network of the improved PANet module and the Faster R-CNN network, and the data is carried out in the data set. It is verified that the mAP of the improved YOLOv4 network is higher than the mAP of the original unimproved YOLOv4 network. Compared with the original network, YOLOv4 is more suitable for student detection and recognition.
{"title":"Student behavior detection based on YOLOv4-Bi","authors":"Xiao-ling Ren, Deyi Yang","doi":"10.1109/CSAIEE54046.2021.9543310","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543310","url":null,"abstract":"With the continuous development of the times, intelligent teaching assistance is also attracting more and more attention in education. The intelligent detection algorithm for student behavior is gradually becoming more precise. In university classrooms, students sleep on mobile phones and are more serious. Intelligent education can more accurately identify student behaviors to help teachers optimize teaching methods, thereby improving students' classroom learning effects. This paper studies and improves YOLOv4, and proposes a network structure called YOLOv4-Bi, which mainly adds the enhanced feature extraction network of YOLOv4 to the feature extraction structure of jumping and top-down, bottom-up combined paths. The used student classroom recording video is enhanced by taking the frame data and training, and testing on this data set. The original YOLOv4 is compared with the network of the improved PANet module and the Faster R-CNN network, and the data is carried out in the data set. It is verified that the mAP of the improved YOLOv4 network is higher than the mAP of the original unimproved YOLOv4 network. Compared with the original network, YOLOv4 is more suitable for student detection and recognition.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"738 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122952410","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543325
Meilin Zhang, Junqiu Li, Qinghui Zhang, Jiale Xu
Nondestructive testing technology of wood acoustic emission(AE) signal is of great significance to evaluate wood internal damage. In order to achieve more accurate and adaptive evaluation, we propose an AE signal analysis method combining instantaneous frequency and power to extract the signal features of different the Intrinsic Mode Function(IMF) components. Then input the SVM classifier for classification and recognition, and adopt the Receiver Operating Characteristic (ROC) curve as the evaluation index to evaluate the classification model of different IMF components. The results show that the instantaneous frequency and power can clearly display AE signal features. The IMF components decomposed by EMD are classified by extracting features, and the classification accuracy of IMF 1 component up to 88% is the highest one. It indicates that IMF 1 component contains a large number of effective AE signal features, which can be utilized for the identification of wood damage and fracture state.
{"title":"Wood acoustic emission signal classification based on IMF's features","authors":"Meilin Zhang, Junqiu Li, Qinghui Zhang, Jiale Xu","doi":"10.1109/CSAIEE54046.2021.9543325","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543325","url":null,"abstract":"Nondestructive testing technology of wood acoustic emission(AE) signal is of great significance to evaluate wood internal damage. In order to achieve more accurate and adaptive evaluation, we propose an AE signal analysis method combining instantaneous frequency and power to extract the signal features of different the Intrinsic Mode Function(IMF) components. Then input the SVM classifier for classification and recognition, and adopt the Receiver Operating Characteristic (ROC) curve as the evaluation index to evaluate the classification model of different IMF components. The results show that the instantaneous frequency and power can clearly display AE signal features. The IMF components decomposed by EMD are classified by extracting features, and the classification accuracy of IMF 1 component up to 88% is the highest one. It indicates that IMF 1 component contains a large number of effective AE signal features, which can be utilized for the identification of wood damage and fracture state.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"47 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129524782","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543344
Yu Liu
Recent works on blind face restoration mainly focus on reference-based methods, which made great progress in recovering high-frequency details and realistic texture from the real world low-quality (LQ) images. However, the multi-scale trait of LQ images is not fully utilized with these methods. Extra face reference also takes up much resources and brings redundant model parameters. In this paper, we introduce the face restoration network with feature prior (FP-FRN) consisting of an adversarial network with a multi-scale feature extraction network which utilizes the multi-scale facial feature to preserve low-level facial characteristics and predict high-level details. Compared to other state-of-the-art approaches, i.e., DFDNet, PSFR-GAN, out FP-FRN generates more realistic texture details and better preserved the low-level feature of the LQ images such as color and shape. As demonstrated by experiments on datasets of synthesized and real LQ images, FP-FRN is superior over other methods.
{"title":"Face Restoration Network with Feature Prior","authors":"Yu Liu","doi":"10.1109/CSAIEE54046.2021.9543344","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543344","url":null,"abstract":"Recent works on blind face restoration mainly focus on reference-based methods, which made great progress in recovering high-frequency details and realistic texture from the real world low-quality (LQ) images. However, the multi-scale trait of LQ images is not fully utilized with these methods. Extra face reference also takes up much resources and brings redundant model parameters. In this paper, we introduce the face restoration network with feature prior (FP-FRN) consisting of an adversarial network with a multi-scale feature extraction network which utilizes the multi-scale facial feature to preserve low-level facial characteristics and predict high-level details. Compared to other state-of-the-art approaches, i.e., DFDNet, PSFR-GAN, out FP-FRN generates more realistic texture details and better preserved the low-level feature of the LQ images such as color and shape. As demonstrated by experiments on datasets of synthesized and real LQ images, FP-FRN is superior over other methods.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"84 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"124629688","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543229
Jingshuo Liu, Shu Zhang, Xinrui Ma, Maoxuan Feng
Image2image translation is one of the most popular image processing tasks. In this work, we use the powerful CycleGAN model and some traditional image processing technology to transform images of cloud into sketch portraits of specific objects. Precisely, this work extract the outer contour of the cloud and use the trained CycleGAN model to transform the outer contour into a specific image (taking the sketch of fish as an example) and the output of the model shows its good translation effect. Moreover, this work set the images of cloud without contour extraction as the control group, which proves the necessity of our preprocessing technology.
{"title":"CycleGAN -based Cloud2painting Translation","authors":"Jingshuo Liu, Shu Zhang, Xinrui Ma, Maoxuan Feng","doi":"10.1109/CSAIEE54046.2021.9543229","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543229","url":null,"abstract":"Image2image translation is one of the most popular image processing tasks. In this work, we use the powerful CycleGAN model and some traditional image processing technology to transform images of cloud into sketch portraits of specific objects. Precisely, this work extract the outer contour of the cloud and use the trained CycleGAN model to transform the outer contour into a specific image (taking the sketch of fish as an example) and the output of the model shows its good translation effect. Moreover, this work set the images of cloud without contour extraction as the control group, which proves the necessity of our preprocessing technology.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"31 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126840275","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543453
Jialei Song, Lijun Jin, Yingpeng Xie, Congmou Wei
To address the problem that the difficulty of selecting parameters in the XGBoost model makes it difficult to optimize the regression effect, a short-term load forecasting model based on the sparrow search algorithm to optimize XGBoost is proposed. Similar days are selected as the training set by the GRA algorithm, the mean absolute error obtained by cross-validation is used as the fitness function, the sparrow search algorithm (SSA) is used to optimize the XGBoost covariate selection process, and the SSA-XGBoost load forecasting model is constructed, and finally the load is corrected by the compensation model to obtain the final load forecasting data. Taking the load data of a region in Zhejiang Province from January 2019 to December 2020 as an example, the prediction ability of the SSA-XGBoost load forecasting model is examined through five experiments. The experimental results show that (i) the parameters of SVM, RF, and XGBoost models can be optimized using the SSA algorithm, and SSA-SVM, SSA-RF, and SSA-XGBoost can quickly calculate the load forecasting data, among which the SSA-XGBoost model has the highest accuracy. Compared with kmeans and other clustering methods, this paper uses the GRA algorithm to select similar days more reasonably, with smaller prediction errors and a controllable number of training sets. The compensation model improves the prediction accuracy of the model by correcting the SSA-XGBoost load prediction data.
{"title":"Optimized XGBoost based sparrow search algorithm for short-term load forecasting","authors":"Jialei Song, Lijun Jin, Yingpeng Xie, Congmou Wei","doi":"10.1109/CSAIEE54046.2021.9543453","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543453","url":null,"abstract":"To address the problem that the difficulty of selecting parameters in the XGBoost model makes it difficult to optimize the regression effect, a short-term load forecasting model based on the sparrow search algorithm to optimize XGBoost is proposed. Similar days are selected as the training set by the GRA algorithm, the mean absolute error obtained by cross-validation is used as the fitness function, the sparrow search algorithm (SSA) is used to optimize the XGBoost covariate selection process, and the SSA-XGBoost load forecasting model is constructed, and finally the load is corrected by the compensation model to obtain the final load forecasting data. Taking the load data of a region in Zhejiang Province from January 2019 to December 2020 as an example, the prediction ability of the SSA-XGBoost load forecasting model is examined through five experiments. The experimental results show that (i) the parameters of SVM, RF, and XGBoost models can be optimized using the SSA algorithm, and SSA-SVM, SSA-RF, and SSA-XGBoost can quickly calculate the load forecasting data, among which the SSA-XGBoost model has the highest accuracy. Compared with kmeans and other clustering methods, this paper uses the GRA algorithm to select similar days more reasonably, with smaller prediction errors and a controllable number of training sets. The compensation model improves the prediction accuracy of the model by correcting the SSA-XGBoost load prediction data.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"104 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134449703","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543137
Taifen Bao, Huimin Jiao, Su Gao, Jifei Cai, Yuansheng Qi
N owadays, medical plastic gloves are sorted into the left and the right hands manually with low efficiency during productive process. In this paper, an automated way is proposed to improve this situation through establishing a convolutional neural network model for image recognition. The back propagation process of learning and training is analyzed in order to optimize the weight by adopting the combination of different activation layers and different loss functions. For the same learning times, there are two evaluation indexes. One is the result of recognition accuracy in the training set, the other is the convergence curve and oscillation amplitude of the loss function. Finally, the adaptability of the combinations is discussed, which plays an important role in improving the recognition accuracy of the left and the right hand.
{"title":"Back Propagation Optimization of Convolutional Neural Network Based on the left and the right hands Identification","authors":"Taifen Bao, Huimin Jiao, Su Gao, Jifei Cai, Yuansheng Qi","doi":"10.1109/CSAIEE54046.2021.9543137","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543137","url":null,"abstract":"N owadays, medical plastic gloves are sorted into the left and the right hands manually with low efficiency during productive process. In this paper, an automated way is proposed to improve this situation through establishing a convolutional neural network model for image recognition. The back propagation process of learning and training is analyzed in order to optimize the weight by adopting the combination of different activation layers and different loss functions. For the same learning times, there are two evaluation indexes. One is the result of recognition accuracy in the training set, the other is the convergence curve and oscillation amplitude of the loss function. Finally, the adaptability of the combinations is discussed, which plays an important role in improving the recognition accuracy of the left and the right hand.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"46 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115900769","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543328
Yong Li, Chu He, Qile Zhao, Jiarui Hu
GNSS system is one of the most widely used wireless systems. A GNSS receiver must lock on satellite signals effectively and quickly. The fastest known algorithm to solve this problem is based on the Fast Fourier Transform(FFT) and the Invert Fast Fourier Transform(IFFF). This paper proposed a novel architecture to implement GNSS signal acquisition system on digital IC or FPGA. Former researchers tended to use FIR filter to compensate for the CIC filter. By adding the CIC filter only, the proposed system aims to reduce the overall calculation complexity by reducing the FFT size. By simplifying the calculation of GNSS acquisition, the memory usage efficiency will be highly improved. Additionally, there will be a significant reduction in GNSS system power consumption.
{"title":"Optimization of GNSS Signals Acquisition Algorithm Complexity Using Comb Decimation Filter","authors":"Yong Li, Chu He, Qile Zhao, Jiarui Hu","doi":"10.1109/CSAIEE54046.2021.9543328","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543328","url":null,"abstract":"GNSS system is one of the most widely used wireless systems. A GNSS receiver must lock on satellite signals effectively and quickly. The fastest known algorithm to solve this problem is based on the Fast Fourier Transform(FFT) and the Invert Fast Fourier Transform(IFFF). This paper proposed a novel architecture to implement GNSS signal acquisition system on digital IC or FPGA. Former researchers tended to use FIR filter to compensate for the CIC filter. By adding the CIC filter only, the proposed system aims to reduce the overall calculation complexity by reducing the FFT size. By simplifying the calculation of GNSS acquisition, the memory usage efficiency will be highly improved. Additionally, there will be a significant reduction in GNSS system power consumption.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114358891","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 : 2021-08-20DOI: 10.1109/CSAIEE54046.2021.9543192
Haoqian Xue, Keyu Ren
Convolutional neural network (CNN) is the main tool for deep learning and computer vision, and it has many applications in face recognition, sign language recognition and speech recognition. As deep learning becomes more and more mature, the application of convolutional neural networks will become more and more widespread. As we know, the deeper a neural network is, the higher its memory and computational power overhead. Many neural networks used in medicine, autonomous driving, and language recognition have large model complexity, which makes it difficult to apply these CNNs to people's daily life. Therefore, the development of simple, lightweight and small neural networks has become the focus of researchers nowadays. In this paper, we summarize the development of convolutional neural networks in recent years, as well as various methods for compressing models and migrating data from large models to small ones. In general, the main convolutional neural network compression approaches are: pruning, knowledge distillation, aggregating neurons of different scales, proposing new structures, etc. We start from the structure of neural networks, introduce the major structural changes experienced from the development of convolutional neural networks, and then analyze various pruning, compression and knowledge distillation methods. For specific methods, we run different models and compare the improvements of the new methods with respect to the old ones. We also debugged models on adversarial generative pruning, teacher-student networks, and other compressed CNNs during this period, and drew some constructive conclusions. Finally, we summarize the trends in CNN development in recent years and the challenges we may face in the future.
{"title":"Recent research trends on Model Compression and Knowledge Transfer in CNNs","authors":"Haoqian Xue, Keyu Ren","doi":"10.1109/CSAIEE54046.2021.9543192","DOIUrl":"https://doi.org/10.1109/CSAIEE54046.2021.9543192","url":null,"abstract":"Convolutional neural network (CNN) is the main tool for deep learning and computer vision, and it has many applications in face recognition, sign language recognition and speech recognition. As deep learning becomes more and more mature, the application of convolutional neural networks will become more and more widespread. As we know, the deeper a neural network is, the higher its memory and computational power overhead. Many neural networks used in medicine, autonomous driving, and language recognition have large model complexity, which makes it difficult to apply these CNNs to people's daily life. Therefore, the development of simple, lightweight and small neural networks has become the focus of researchers nowadays. In this paper, we summarize the development of convolutional neural networks in recent years, as well as various methods for compressing models and migrating data from large models to small ones. In general, the main convolutional neural network compression approaches are: pruning, knowledge distillation, aggregating neurons of different scales, proposing new structures, etc. We start from the structure of neural networks, introduce the major structural changes experienced from the development of convolutional neural networks, and then analyze various pruning, compression and knowledge distillation methods. For specific methods, we run different models and compare the improvements of the new methods with respect to the old ones. We also debugged models on adversarial generative pruning, teacher-student networks, and other compressed CNNs during this period, and drew some constructive conclusions. Finally, we summarize the trends in CNN development in recent years and the challenges we may face in the future.","PeriodicalId":376014,"journal":{"name":"2021 IEEE International Conference on Computer Science, Artificial Intelligence and Electronic Engineering (CSAIEE)","volume":"67 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2021-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126210127","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：