ISSN 
2019 Vol. 48, No. 5
2019, 48(5): 641-641.
Abstract:
2019, 48(5): 642-642.
Abstract:
2019, 48(5): 643-643.
Abstract:
2019, 48(5): 655-673.
doi: 10.3969/j.issn.1001-0548.2019.05.002
Abstract:
In this review, we attempt to highlight major concepts, techniques, challenges and future trends of context-aware recommender systems in social and scientific domains. The primary objective of this paper is to sum up the most recent developments in this rich knowledge area. A set of techniques and major frameworks available for context-based recommender systems are classified and introduced. Along with classical content-based, collaborative filtering and matrix factorization based techniques, we investigate the most recent research areas, i.e., deep learning and fuzzy logic based methodologies. Finally, we close by portraying potential future research opportunities with respect to utilizing context information in recommendation process.
In this review, we attempt to highlight major concepts, techniques, challenges and future trends of context-aware recommender systems in social and scientific domains. The primary objective of this paper is to sum up the most recent developments in this rich knowledge area. A set of techniques and major frameworks available for context-based recommender systems are classified and introduced. Along with classical content-based, collaborative filtering and matrix factorization based techniques, we investigate the most recent research areas, i.e., deep learning and fuzzy logic based methodologies. Finally, we close by portraying potential future research opportunities with respect to utilizing context information in recommendation process.
2019, 48(5): 679-679.
Abstract:
2019, 48(5): 681-681.
Abstract:
2019, 48(5): 644-654.
doi: 10.3969/j.issn.1001-0548.2019.05.001
Abstract:
Human parsing aims at identifying the body parts and clothing items from human images at pixel level. This paper investigates and analyzes the approaches of human parsing based on deep learning, which mainly includes three aspects:the basic technologies involved in human parsing, the main datasets and evaluation standard, and the existing methods. Firstly, the basic technologies involved in human parsing based on deep learning, including convolutional neural network and semantic segmentation are reviewed. Secondly, this paper introduces 8 datasets for human parsing in detail according to the number of images, the number of categories, advantages and disadvantages. In addition, four commonly used evaluation metrics are summarized. Finally, existing representative schemes for human paring based on deep learning are concerned, including feature enhancement, structure of human body, multi-task learning, and generative adversarial networks. This paper summarizes the approaches of instance-level human parsing, and presents some ideas worth studying.
Human parsing aims at identifying the body parts and clothing items from human images at pixel level. This paper investigates and analyzes the approaches of human parsing based on deep learning, which mainly includes three aspects:the basic technologies involved in human parsing, the main datasets and evaluation standard, and the existing methods. Firstly, the basic technologies involved in human parsing based on deep learning, including convolutional neural network and semantic segmentation are reviewed. Secondly, this paper introduces 8 datasets for human parsing in detail according to the number of images, the number of categories, advantages and disadvantages. In addition, four commonly used evaluation metrics are summarized. Finally, existing representative schemes for human paring based on deep learning are concerned, including feature enhancement, structure of human body, multi-task learning, and generative adversarial networks. This paper summarizes the approaches of instance-level human parsing, and presents some ideas worth studying.
2019, 48(5): 674-678.
doi: 10.3969/j.issn.1001-0548.2019.05.003
Abstract:
With the development of Generative Adversarial Networks (GAN), more and more researches have been conducted in the field of Chinese fonts transformation and researchers are able to generate high-quality images of Chinese characters. These font transformation models can transform a source font to a target font using GAN. However, current methods have limitations that 1) generated images are oftentimes blurry and 2) models can only learn and produce one target font at a time. To address these problems, we have developed a brand-new model to perform Chinese font transformation. First, font information is attached to images to tell the generator the fonts that we want to transform. Then, the generator extracts and learns feature mappings through convolutional networks and generates photo-realistic images using transposed convolutional networks. The ground truth images are then used as supervisory information to ensure that characters and fonts generated are consistent with themselves. This model only needs to be trained once, but it is able to transform one font to multiple fonts and produce new fonts. Extensive experiments on seven Chinese font datasets show the superiority of the proposed method over several other methods in Chinese font transformation.
With the development of Generative Adversarial Networks (GAN), more and more researches have been conducted in the field of Chinese fonts transformation and researchers are able to generate high-quality images of Chinese characters. These font transformation models can transform a source font to a target font using GAN. However, current methods have limitations that 1) generated images are oftentimes blurry and 2) models can only learn and produce one target font at a time. To address these problems, we have developed a brand-new model to perform Chinese font transformation. First, font information is attached to images to tell the generator the fonts that we want to transform. Then, the generator extracts and learns feature mappings through convolutional networks and generates photo-realistic images using transposed convolutional networks. The ground truth images are then used as supervisory information to ensure that characters and fonts generated are consistent with themselves. This model only needs to be trained once, but it is able to transform one font to multiple fonts and produce new fonts. Extensive experiments on seven Chinese font datasets show the superiority of the proposed method over several other methods in Chinese font transformation.
Porous Fe/N-doped Carbon Layers Wrapped CNTs Electrocatalysts for Superior Oxygen Reduction Reaction
2019, 48(5): 682-689.
doi: 10.3969/j.issn.1001-0548.2019.05.004
Abstract:
Developing high performance, low-cost oxygen reduction electrocatalysts is the key to reduce the cost of fuel cells. The transition metal-nitrogen-carbon materials have been highlighted as promising candidates owing to their high catalytic activity, low cost and environmental friendliness. Herein, we report a facile polydopamine modification approach to prepare porous CNTs@Fe/N/C elecrtocatalysts via pyrolysis of the CNTs@PDA and FeCl3 composites. The morphology and composition of these as-prepared catalysts were characterized by TEM, BET, Raman and XPS. The electrochemical results show that the half-wave potential of the CNTs@40% Fe/N/C catalyst is as high as 0.881 V, which is close to that of commercial Pt/C catalyst. Moreover, the CNTs@40% Fe/N/C also showed superior durability and resistance to the methanol crossover.
Developing high performance, low-cost oxygen reduction electrocatalysts is the key to reduce the cost of fuel cells. The transition metal-nitrogen-carbon materials have been highlighted as promising candidates owing to their high catalytic activity, low cost and environmental friendliness. Herein, we report a facile polydopamine modification approach to prepare porous CNTs@Fe/N/C elecrtocatalysts via pyrolysis of the CNTs@PDA and FeCl3 composites. The morphology and composition of these as-prepared catalysts were characterized by TEM, BET, Raman and XPS. The electrochemical results show that the half-wave potential of the CNTs@40% Fe/N/C catalyst is as high as 0.881 V, which is close to that of commercial Pt/C catalyst. Moreover, the CNTs@40% Fe/N/C also showed superior durability and resistance to the methanol crossover.
2019, 48(5): 690-697.
doi: 10.3969/j.issn.1001-0548.2019.05.005
Abstract:
Tar in smoking stream is detrimental to human health and has also exerted pressure on environment. In this work, the polyvinyl alcohol (PVA) nanofibers were fabricated using an electrospinning technique and a filter based on the nanofiber film was designed and implemented to eliminate tar from smoke stream. The filtering efficiency was evaluated by the mass gain of film after absorbing the tar. The fiber diameter, thickness of films and the inhaling rate affecting the quantity of tar absorption and the saturation of tar in filter were also investigated. The possible alien substances to the polymer filters were analyzed by X-ray photoelectron spectroscopy. The work demonstrates a novel filter device to abate hazardous matters in cigarette both for human health and a clean environment.
Tar in smoking stream is detrimental to human health and has also exerted pressure on environment. In this work, the polyvinyl alcohol (PVA) nanofibers were fabricated using an electrospinning technique and a filter based on the nanofiber film was designed and implemented to eliminate tar from smoke stream. The filtering efficiency was evaluated by the mass gain of film after absorbing the tar. The fiber diameter, thickness of films and the inhaling rate affecting the quantity of tar absorption and the saturation of tar in filter were also investigated. The possible alien substances to the polymer filters were analyzed by X-ray photoelectron spectroscopy. The work demonstrates a novel filter device to abate hazardous matters in cigarette both for human health and a clean environment.
2019, 48(5): 698-705.
doi: 10.3969/j.issn.1001-0548.2019.05.006
Abstract:
Single station localization plays an important role in indoor localization. Most of the indoor localization algorithms are combined with multiple Wi-Fi access points (APs) to locate, the localization accuracy is low or even unable to locate when there is only a single AP. In terms of the ubiquitous single AP localization situation, the channel frequency response (CFR) of the commercial three-antenna Wi-Fi network card is used to estimate the angle of arrival (AOA). The CFR amplitude information is applied to estimate the distance from the signal propagation model. Further, we presents a direct path identification algorithm based on the two-dimensional clustering information of AOA and time of arrival (TOA) and use the existing Wi-Fi device with three antennas to measure the angle, distance and localization in the indoor room. The experimental results show that the proposed localization system can achieve a median error of 1.3 m, and can meet the needs of indoor positioning.
Single station localization plays an important role in indoor localization. Most of the indoor localization algorithms are combined with multiple Wi-Fi access points (APs) to locate, the localization accuracy is low or even unable to locate when there is only a single AP. In terms of the ubiquitous single AP localization situation, the channel frequency response (CFR) of the commercial three-antenna Wi-Fi network card is used to estimate the angle of arrival (AOA). The CFR amplitude information is applied to estimate the distance from the signal propagation model. Further, we presents a direct path identification algorithm based on the two-dimensional clustering information of AOA and time of arrival (TOA) and use the existing Wi-Fi device with three antennas to measure the angle, distance and localization in the indoor room. The experimental results show that the proposed localization system can achieve a median error of 1.3 m, and can meet the needs of indoor positioning.
2019, 48(5): 706-710.
doi: 10.3969/j.issn.1001-0548.2019.05.007
Abstract:
The existing algorithm against deception-false-target by homogeneous radar system is single due to the simplicity of configured mode. To solve the problem, a method against deception-false-target based on isomerous netted radars is proposed. The isomerous netted radars in this paper consists of 3D radar and 2D radar. The distance and azimuth information between target and 2D radar are estimated by making use of the measurement of 3D radar. By comparing the estimated values with the measurements of 2D radar, the false-target can be discriminated by using chi-square statistics. We further analyze the influence of the distance of deception distance, observation frequency, the 2D radar accuracy on algorithm's discrimination probability. Simulation results show that the proposed method is efficacious in discriminating the false-target jamming.
The existing algorithm against deception-false-target by homogeneous radar system is single due to the simplicity of configured mode. To solve the problem, a method against deception-false-target based on isomerous netted radars is proposed. The isomerous netted radars in this paper consists of 3D radar and 2D radar. The distance and azimuth information between target and 2D radar are estimated by making use of the measurement of 3D radar. By comparing the estimated values with the measurements of 2D radar, the false-target can be discriminated by using chi-square statistics. We further analyze the influence of the distance of deception distance, observation frequency, the 2D radar accuracy on algorithm's discrimination probability. Simulation results show that the proposed method is efficacious in discriminating the false-target jamming.
2019, 48(5): 711-716.
doi: 10.3969/j.issn.1001-0548.2019.05.008
Abstract:
Grid cell is an important neuron cell in the navigation loop of animal brain. It can complete the function of path integration autonomously, and can complete complex environment space navigation with other brain regions. In this paper, the reaction-diffusion mechanism is used to model the formation of grid-cell firing field in autonomous navigation. The synaptic connections between the grid-cell groups, the nonlinear restriction of the grid-cells firing rate, and the response of the excitatory grid-cells to the ego-emotion information are obtained respectively. The simulation results show that the proposed model can effectively simulate the formation process of grid-like firing rate distribution in the grid-cell groups, and the formation process of hexagon firing field of individual grid-cells based on the ego-motion information. The proposed model can provide a reference for the construction of brain-inspired navigation system.
Grid cell is an important neuron cell in the navigation loop of animal brain. It can complete the function of path integration autonomously, and can complete complex environment space navigation with other brain regions. In this paper, the reaction-diffusion mechanism is used to model the formation of grid-cell firing field in autonomous navigation. The synaptic connections between the grid-cell groups, the nonlinear restriction of the grid-cells firing rate, and the response of the excitatory grid-cells to the ego-emotion information are obtained respectively. The simulation results show that the proposed model can effectively simulate the formation process of grid-like firing rate distribution in the grid-cell groups, and the formation process of hexagon firing field of individual grid-cells based on the ego-motion information. The proposed model can provide a reference for the construction of brain-inspired navigation system.
2019, 48(5): 717-721.
doi: 10.3969/j.issn.1001-0548.2019.05.009
Abstract:
Urban mapping with a great level of details becomes possible using the decimeter-level resolution synthetic aperture radar (SAR) data. With available interferometric synthetic aperture radar (InSAR) techniques, it is possible to retrieve the 3D information of individual buildings. In China, the high-rise buildings are numerous in urban areas. It is of the great importance to obtain the 3D information of these high-rise buildings in the city planning. In this study, we show the possibility of the high-rise building 3D reconstruction using only one wrapped InSAR phase imagery with high-resolution. The effectiveness of the approach including the building detection, facade extraction, and 3D information estimation has been demonstrated with the acquired airborne N-SAR data.
Urban mapping with a great level of details becomes possible using the decimeter-level resolution synthetic aperture radar (SAR) data. With available interferometric synthetic aperture radar (InSAR) techniques, it is possible to retrieve the 3D information of individual buildings. In China, the high-rise buildings are numerous in urban areas. It is of the great importance to obtain the 3D information of these high-rise buildings in the city planning. In this study, we show the possibility of the high-rise building 3D reconstruction using only one wrapped InSAR phase imagery with high-resolution. The effectiveness of the approach including the building detection, facade extraction, and 3D information estimation has been demonstrated with the acquired airborne N-SAR data.
2019, 48(5): 722-727.
doi: 10.3969/j.issn.1001-0548.2019.05.010
Abstract:
In this paper, an one time programmable (OTP) memory controller circuit is designed. By operating the OTP memory instruction, the interface timings corresponding to different operations are generated to complete operations such as programming, reading, waking up, resetting, and sleeping on the OTP memory. At the same time, in the OTP memory programming operation, a programming algorithm is designed for the OTP programming error-prone problem. That is, the OTP programming address is redundantly processed, and the pulse voltage is applied to the same address multiple times during the programming operation. The accuracy of data programmed into the OTP memory is dramatically increased. The embedded programming algorithm circuit implements a highly reliable programming algorithm and effectively controls the programming operation of the OTP memory. Therefore, it solved the problem of programming error when accessing memory, and greatly improved the reliability of the OTP memory.
In this paper, an one time programmable (OTP) memory controller circuit is designed. By operating the OTP memory instruction, the interface timings corresponding to different operations are generated to complete operations such as programming, reading, waking up, resetting, and sleeping on the OTP memory. At the same time, in the OTP memory programming operation, a programming algorithm is designed for the OTP programming error-prone problem. That is, the OTP programming address is redundantly processed, and the pulse voltage is applied to the same address multiple times during the programming operation. The accuracy of data programmed into the OTP memory is dramatically increased. The embedded programming algorithm circuit implements a highly reliable programming algorithm and effectively controls the programming operation of the OTP memory. Therefore, it solved the problem of programming error when accessing memory, and greatly improved the reliability of the OTP memory.
2019, 48(5): 728-733.
doi: 10.3969/j.issn.1001-0548.2019.05.011
Abstract:
The research on the influence factors and recognition of electrical fast transient burst (EFT/B) is of great significance to the accuracy of relay protection action in power system. For the phenomenon of malfunction caused by EFT/B in the transformer neutral point gap protection, this paper studies its impact on the transformer neutral point overvoltage. Based on the existing segmented arc model, a simplified model is proposed to reduce the difficulty of establishing the traditional simulation model. On the basis of the analysis of the field relationship of breaker, a new method is proposed to determine the alternating current (AC) arc. In order to study the circuit breaker breaking at different times, breaking of different short-circuit types, breaking at different speeds on the impact transformer neutral point over-voltage size, this paper establishes a circuit breaker model in MATLAB. The time-frequency characteristics and energy distribution characteristics of different transformer neutral point overvoltages caused by EFT/B, lightning and single-phase grounding respectively are generated by using wavelet decomposition. Further, a criterion is set up according to the high and low frequency energy ratio. The simulation results show that the criterion can distinguish the three kinds of overvoltages accurately.
The research on the influence factors and recognition of electrical fast transient burst (EFT/B) is of great significance to the accuracy of relay protection action in power system. For the phenomenon of malfunction caused by EFT/B in the transformer neutral point gap protection, this paper studies its impact on the transformer neutral point overvoltage. Based on the existing segmented arc model, a simplified model is proposed to reduce the difficulty of establishing the traditional simulation model. On the basis of the analysis of the field relationship of breaker, a new method is proposed to determine the alternating current (AC) arc. In order to study the circuit breaker breaking at different times, breaking of different short-circuit types, breaking at different speeds on the impact transformer neutral point over-voltage size, this paper establishes a circuit breaker model in MATLAB. The time-frequency characteristics and energy distribution characteristics of different transformer neutral point overvoltages caused by EFT/B, lightning and single-phase grounding respectively are generated by using wavelet decomposition. Further, a criterion is set up according to the high and low frequency energy ratio. The simulation results show that the criterion can distinguish the three kinds of overvoltages accurately.
2019, 48(5): 734-740.
doi: 10.3969/j.issn.1001-0548.2019.05.012
Abstract:
Analog realization error and channel mismatch error comprised of offset error, gain error and time error are the two key kinds of error sources in frequency interleaved analog to digital converter (ADC) systems, resulting in the systems' global performance degeneration. In this paper, intensive study is carried out on the errors in frequency interleaved ADC systems. On the one hand, it shows that the virtual frequency response of the sampling filter leading to the realization error is mathematically calculable. On the other hand, it also reveals that the channel mismatch error will give birth to the regular appearance of the spurious noise peaks in the spectrum of the output signal. Both the theoretical derivation and simulation verification certify that the spurious noises caused by the gain error and the time error occur at the same frequency location but not that of phase, and the peak amplitude caused by the offset error has nothing to do with the input signal frequency. The research results will promote the development of error compensation and calibration.
Analog realization error and channel mismatch error comprised of offset error, gain error and time error are the two key kinds of error sources in frequency interleaved analog to digital converter (ADC) systems, resulting in the systems' global performance degeneration. In this paper, intensive study is carried out on the errors in frequency interleaved ADC systems. On the one hand, it shows that the virtual frequency response of the sampling filter leading to the realization error is mathematically calculable. On the other hand, it also reveals that the channel mismatch error will give birth to the regular appearance of the spurious noise peaks in the spectrum of the output signal. Both the theoretical derivation and simulation verification certify that the spurious noises caused by the gain error and the time error occur at the same frequency location but not that of phase, and the peak amplitude caused by the offset error has nothing to do with the input signal frequency. The research results will promote the development of error compensation and calibration.
2019, 48(5): 741-746.
doi: 10.3969/j.issn.1001-0548.2019.05.013
Abstract:
In non-destructive evaluation area, defect feature extraction and analysis based on eddy current pulsed thermography (ECPT) technique is a research focus. In this paper, a novel defect feature extraction approach is proposed to highlight the defect information in ECPT. The proposed approach includes entropy-based image selection, local (element-wise) sparse decomposition and image fusion. Comparing with other two common feature extraction algorithms, independent component analysis and robust principal component analysis, the proposed algorithm can extract more defect features and suppress background.
In non-destructive evaluation area, defect feature extraction and analysis based on eddy current pulsed thermography (ECPT) technique is a research focus. In this paper, a novel defect feature extraction approach is proposed to highlight the defect information in ECPT. The proposed approach includes entropy-based image selection, local (element-wise) sparse decomposition and image fusion. Comparing with other two common feature extraction algorithms, independent component analysis and robust principal component analysis, the proposed algorithm can extract more defect features and suppress background.
2019, 48(5): 747-753.
doi: 10.3969/j.issn.1001-0548.2019.05.014
Abstract:
For hand pose estimation, one immediate problem is to reduce the need for labeled data which is difficult to provide in desired quantity, realism and accuracy. To meet this need, a novel multi-view projection based semi-supervised learning algorithm is proposed. Firstly, 3D hand points are extracted from a single depth image without label and projected onto three orthogonal planes. Secondly, an encoder-decoder model is applied to learn the latent representation of two projections. Finally, small amount of labeled data is used to learn a mapping from latent representation to hand joint coordinates. The propose algorithm is evaluated on NYU hand pose estimation dataset, and the experimental results demonstrate the effectiveness and advantages of our proposed algorithm.
For hand pose estimation, one immediate problem is to reduce the need for labeled data which is difficult to provide in desired quantity, realism and accuracy. To meet this need, a novel multi-view projection based semi-supervised learning algorithm is proposed. Firstly, 3D hand points are extracted from a single depth image without label and projected onto three orthogonal planes. Secondly, an encoder-decoder model is applied to learn the latent representation of two projections. Finally, small amount of labeled data is used to learn a mapping from latent representation to hand joint coordinates. The propose algorithm is evaluated on NYU hand pose estimation dataset, and the experimental results demonstrate the effectiveness and advantages of our proposed algorithm.
2019, 48(5): 754-758.
doi: 10.3969/j.issn.1001-0548.2019.05.015
Abstract:
Intelligent monitoring ball machine is widely used in indoor intelligent monitoring. Aiming at the problem of long-term detection, recognition and tracking of intrusive targets, this paper designs and implements a closed-loop structure combining target detection, target recognition and target tracking algorithm, and uses control algorithm to control the ball machine platform to automatically follow the intrusive targets. Vibe algorithm is used for moving target detection and neural network is applied for target recognition, where single shot multi box detector (SSD) network is used to detect face and yolov3 (You only look once) network is used to recognize human body. After recognizing the tracking target, discriminative correlation filter with channel and spatial reliability (csr-dcf) target tracking algorithm is used to track the target. In the tracking mode, the fuzzy pid control algorithm is started to control the platform to follow the target rotation, and after locking the target, the tracking mode is used to track the target. The model is switched to target recognition mode again, forming a closed loop of detection, recognition, tracking and control. The test shows that this method improves the robustness of intrusion tracking function, and can be tracked for a long time in the case of fast movement of the intrusive target, occlusion and temporary disappearance.
Intelligent monitoring ball machine is widely used in indoor intelligent monitoring. Aiming at the problem of long-term detection, recognition and tracking of intrusive targets, this paper designs and implements a closed-loop structure combining target detection, target recognition and target tracking algorithm, and uses control algorithm to control the ball machine platform to automatically follow the intrusive targets. Vibe algorithm is used for moving target detection and neural network is applied for target recognition, where single shot multi box detector (SSD) network is used to detect face and yolov3 (You only look once) network is used to recognize human body. After recognizing the tracking target, discriminative correlation filter with channel and spatial reliability (csr-dcf) target tracking algorithm is used to track the target. In the tracking mode, the fuzzy pid control algorithm is started to control the platform to follow the target rotation, and after locking the target, the tracking mode is used to track the target. The model is switched to target recognition mode again, forming a closed loop of detection, recognition, tracking and control. The test shows that this method improves the robustness of intrusion tracking function, and can be tracked for a long time in the case of fast movement of the intrusive target, occlusion and temporary disappearance.
2019, 48(5): 759-766.
doi: 10.3969/j.issn.1001-0548.2019.05.016
Abstract:
Text sentiment classification is a hot topic in the field of natural language processing in recent years. It aims to analyze the subjective sentiment polarity of text. More and more attention has been paid to the problem of fine grained sentiment classification based on specific aspects. In traditional deep models, the attention mechanism can significantly improve the classification performance. Based on the characteristics of Chinese language, a deep model combining multi-hop attention mechanism and convolutional neural network (MHA-CNN) is proposed. The model makes use of the multidimensional combination features to remedy the deficiency of one dimensional feature attention mechanism, and can get deeper aspect sentiment feature information without any prior knowledge. Relative to the attention mechanism based long short-term memory (LSTM) network, the model has smaller time overhead and can retain word order information of the characteristic part. Finally, we conduct experiments on a network open Chinese data set (including 6 kinds of field data), and get better classification results than the ordinary deep network model, the attention-based LSTM model and the attention-based deep memory network model.
Text sentiment classification is a hot topic in the field of natural language processing in recent years. It aims to analyze the subjective sentiment polarity of text. More and more attention has been paid to the problem of fine grained sentiment classification based on specific aspects. In traditional deep models, the attention mechanism can significantly improve the classification performance. Based on the characteristics of Chinese language, a deep model combining multi-hop attention mechanism and convolutional neural network (MHA-CNN) is proposed. The model makes use of the multidimensional combination features to remedy the deficiency of one dimensional feature attention mechanism, and can get deeper aspect sentiment feature information without any prior knowledge. Relative to the attention mechanism based long short-term memory (LSTM) network, the model has smaller time overhead and can retain word order information of the characteristic part. Finally, we conduct experiments on a network open Chinese data set (including 6 kinds of field data), and get better classification results than the ordinary deep network model, the attention-based LSTM model and the attention-based deep memory network model.
2019, 48(5): 767-773.
doi: 10.3969/j.issn.1001-0548.2019.05.017
Abstract:
Recently, dictionary learning (DL) has been applied to various pattern recognition tasks successfully, analysis dictionary learning, however as an important branch of dictionary learning, has not been fully exploited due to its poor discriminability. In this paper, a novel robust and discriminative analysis dictionary learning method is proposed, which specially seeks low rank representation from noisy data and learn a discriminative dictionary from the recovered clean data by incorporating with the Fisher criterion. The discriminability of dictionary is improved by introducing the supervised mechanism. At last, the task of human action recognition is conducted by applying the proposed method. Experiments on several human action recognition datasets show that the proposed method outperforms other classical synthesis dictionary methods.
Recently, dictionary learning (DL) has been applied to various pattern recognition tasks successfully, analysis dictionary learning, however as an important branch of dictionary learning, has not been fully exploited due to its poor discriminability. In this paper, a novel robust and discriminative analysis dictionary learning method is proposed, which specially seeks low rank representation from noisy data and learn a discriminative dictionary from the recovered clean data by incorporating with the Fisher criterion. The discriminability of dictionary is improved by introducing the supervised mechanism. At last, the task of human action recognition is conducted by applying the proposed method. Experiments on several human action recognition datasets show that the proposed method outperforms other classical synthesis dictionary methods.
2019, 48(5): 774-778.
doi: 10.3969/j.issn.1001-0548.2019.05.018
Abstract:
Based on the convolutional neural networks (CNNs) model, an image classification method of model fusing is proposed. The original data is composed of enhanced images and normalized data, and the mapping data is generated by negating original data. Then the CNNs models with the original data and the mapping data are trained separately. Next the two sets of CNNs models are fused to obtain the improved of CNNs model after training. The improved method is generalized to some more complex CNNs models after it is proved effective for simple cases through hypothesis, verification, and theoretical derivation steps. The experimental results show that the model after the fusion performs well. Compared with the original CNNs model, the classification accuracy the proposed model is increased by 1% and 3% based on the sets of CIFAR-10 and CIFAR-100 data, respectively.
Based on the convolutional neural networks (CNNs) model, an image classification method of model fusing is proposed. The original data is composed of enhanced images and normalized data, and the mapping data is generated by negating original data. Then the CNNs models with the original data and the mapping data are trained separately. Next the two sets of CNNs models are fused to obtain the improved of CNNs model after training. The improved method is generalized to some more complex CNNs models after it is proved effective for simple cases through hypothesis, verification, and theoretical derivation steps. The experimental results show that the model after the fusion performs well. Compared with the original CNNs model, the classification accuracy the proposed model is increased by 1% and 3% based on the sets of CIFAR-10 and CIFAR-100 data, respectively.
2019, 48(5): 779-785.
doi: 10.3969/j.issn.1001-0548.2019.05.019
Abstract:
The internal user security behavior assessment method affects the accuracy of the user's operational behavior assessment due to less considers the contextual relevance of the user's operational behaviors. In view of this situation, and considering the characteristics of long-short term memory (LSTM) is suitable for dealing with time series problems, an internal user security behavior evaluation method based on LSTM is proposed. In this method, the data are vectorized firstly and then divided according to the N vs. 1 scheme. The LSTM algorithm is used to uniformly model the known user's behavior habits. Finally, the decision threshold is determined by the bimodal threshold mechanism and user behaviors are evaluated. Experimental results show that the data partitioning scheme of this method improves the ability to detect abnormal operation of unknown users, and by introducing a bimodal threshold mechanism, the accuracy and recall of the algorithm for detecting abnormal operations of unknown users are improved.
The internal user security behavior assessment method affects the accuracy of the user's operational behavior assessment due to less considers the contextual relevance of the user's operational behaviors. In view of this situation, and considering the characteristics of long-short term memory (LSTM) is suitable for dealing with time series problems, an internal user security behavior evaluation method based on LSTM is proposed. In this method, the data are vectorized firstly and then divided according to the N vs. 1 scheme. The LSTM algorithm is used to uniformly model the known user's behavior habits. Finally, the decision threshold is determined by the bimodal threshold mechanism and user behaviors are evaluated. Experimental results show that the data partitioning scheme of this method improves the ability to detect abnormal operation of unknown users, and by introducing a bimodal threshold mechanism, the accuracy and recall of the algorithm for detecting abnormal operations of unknown users are improved.
2019, 48(5): 786-793.
doi: 10.3969/j.issn.1001-0548.2019.05.020
Abstract:
This paper uses the method of regular-triangle mapping to empirically analyze the collaboration between scientists. Based on the data of scientific collaboration networks, each scientist is mapped to a standard regular triangle based on the collaboration relationship with three nearest higher-influence scientists who satisfy a type of given condition. The modes of their relationship can be shown by the pattern of their mapping position in the regular triangle. We find that the distribution of scientists on the mapping triangle shows obvious center-gathering tendency, indicating the strong equilibrium property that the collaboration of scientists is often evenly distributed among multiple collaboration partners with higher-influence. This property can be observed in the full parameter space of all of the four research topics discussed in this paper, and the academic age, the difference in total citations and some other factors do not efficiently impact the expression of the equilibrium property. Furthermore, we find that the impact of the equilibrium property on the scientific influence of scientists mainly is negative. This research provides a new perspective for analyzing the pattern of collaboration relationships between scientists and the node relationships of various complex networks.
This paper uses the method of regular-triangle mapping to empirically analyze the collaboration between scientists. Based on the data of scientific collaboration networks, each scientist is mapped to a standard regular triangle based on the collaboration relationship with three nearest higher-influence scientists who satisfy a type of given condition. The modes of their relationship can be shown by the pattern of their mapping position in the regular triangle. We find that the distribution of scientists on the mapping triangle shows obvious center-gathering tendency, indicating the strong equilibrium property that the collaboration of scientists is often evenly distributed among multiple collaboration partners with higher-influence. This property can be observed in the full parameter space of all of the four research topics discussed in this paper, and the academic age, the difference in total citations and some other factors do not efficiently impact the expression of the equilibrium property. Furthermore, we find that the impact of the equilibrium property on the scientific influence of scientists mainly is negative. This research provides a new perspective for analyzing the pattern of collaboration relationships between scientists and the node relationships of various complex networks.
2019, 48(5): 794-800.
doi: 10.3969/j.issn.1001-0548.2019.05.021
Abstract:
To uncover the networks' structure according to limited interactional information is one of the significant problems in the field of network science. We develop a method for reconstructing and analyzing the networks' structure based on single structure attribute of network. First, a series of synthetic networks with tunable clustering coefficient are generated under Holme-Kim model. Then, the networks' structure is reconstructed by virtue of a compressive-sensing identification model with limited information. Experimental results demonstrate that when we have 20% time-series information of interactions between nodes in the networks, both the average identification accuracy and the average recall of existent links of the whole networks would be increased with the increment of the average clustering coefficient. In this paper, the average clustering coefficient of the target networks is varying from 0.1 to 0.6. The average identification accuracy and the average recall of existent links would reach the optimum value when the network's average clustering coefficient is 0.6. Further, we make a deeper investigation into the experimental data. We find that the average identification accuracy of the networks largely depends on that of the corresponding nodes, whose degree is less than 8 in networks.
To uncover the networks' structure according to limited interactional information is one of the significant problems in the field of network science. We develop a method for reconstructing and analyzing the networks' structure based on single structure attribute of network. First, a series of synthetic networks with tunable clustering coefficient are generated under Holme-Kim model. Then, the networks' structure is reconstructed by virtue of a compressive-sensing identification model with limited information. Experimental results demonstrate that when we have 20% time-series information of interactions between nodes in the networks, both the average identification accuracy and the average recall of existent links of the whole networks would be increased with the increment of the average clustering coefficient. In this paper, the average clustering coefficient of the target networks is varying from 0.1 to 0.6. The average identification accuracy and the average recall of existent links would reach the optimum value when the network's average clustering coefficient is 0.6. Further, we make a deeper investigation into the experimental data. We find that the average identification accuracy of the networks largely depends on that of the corresponding nodes, whose degree is less than 8 in networks.
