2018 Vol. 55 No. 12
2018, 55(12): 2587-2599.
DOI: 10.7544/issn1000-1239.2018.20180623
Abstract：
With the advent of the medical big data era, knowledge interconnection has received extensive attention. How to extract useful medical knowledge from massive data is the key for medical big data analysis. Knowledge graph technology provides a means to extract structured knowledge from massive texts and images.The combination of knowledge graph, big data technology and deep learning technology is becoming the core driving force for the development of artificial intelligence. The knowledge graph technology has a broad application prospect in the medical domain. The application of knowledge graph technology in the medical domain will play an important role in solving the contradiction between the supply of high-quality medical resources and the continuous increase of demand for medical services.At present, the research on medical knowledge graph is still in the exploratory stage. The existing knowledge graph technology generally has several problems such as low efficiency, multiple restrictions and poor expansion in the medical domain. This paper firstly analyzes the medical knowledge graph architecture and construction technology for the strong professionalism and complex structure of big data in the medical domain. Secondly, the key technologies and research progress of the three modules of knowledge extraction, knowledge expression, knowledge fusion and knowledge reasoning in medical knowledge map are summarized. In addition, the application status of medical knowledge maps in clinical decision support, medical intelligence semantic retrieval, medical question answering system and other medical services are introduced. Finally, the existing problems and challenges of current research are discussed and analyzed, and its development is prospected.
With the advent of the medical big data era, knowledge interconnection has received extensive attention. How to extract useful medical knowledge from massive data is the key for medical big data analysis. Knowledge graph technology provides a means to extract structured knowledge from massive texts and images.The combination of knowledge graph, big data technology and deep learning technology is becoming the core driving force for the development of artificial intelligence. The knowledge graph technology has a broad application prospect in the medical domain. The application of knowledge graph technology in the medical domain will play an important role in solving the contradiction between the supply of high-quality medical resources and the continuous increase of demand for medical services.At present, the research on medical knowledge graph is still in the exploratory stage. The existing knowledge graph technology generally has several problems such as low efficiency, multiple restrictions and poor expansion in the medical domain. This paper firstly analyzes the medical knowledge graph architecture and construction technology for the strong professionalism and complex structure of big data in the medical domain. Secondly, the key technologies and research progress of the three modules of knowledge extraction, knowledge expression, knowledge fusion and knowledge reasoning in medical knowledge map are summarized. In addition, the application status of medical knowledge maps in clinical decision support, medical intelligence semantic retrieval, medical question answering system and other medical services are introduced. Finally, the existing problems and challenges of current research are discussed and analyzed, and its development is prospected.
2018, 55(12): 2600-2610.
DOI: 10.7544/issn1000-1239.2018.20180606
Abstract：
Many fragmentation information is highly dispersed in different data sources, such as text, image, video and Web. They are characterized by structural disorder and content one-sided. Current researches implement the extraction, expression and understanding of multi-modal fragmentation information by constructing visual question answering (VQA) system. The VQA task is required to provide the correct answer to a given problem with a corresponding image. The aim of this paper is to design a complete framework and algorithm for image fragmentation information question answering under the basic background of visual question answering task. The main research includes image feature extraction, question text feature extraction, multi-modal feature fusion and answer reasoning. Deep neural network is constructed to extract features for representing images and problem information. Attention mechanism and variational inference method are combined to fusion two modal features of image and problem and reason answers. Experiment results show that the model can effectively extract and understand multi-modal fragmentation information, and improve the accuracy of VQA.
Many fragmentation information is highly dispersed in different data sources, such as text, image, video and Web. They are characterized by structural disorder and content one-sided. Current researches implement the extraction, expression and understanding of multi-modal fragmentation information by constructing visual question answering (VQA) system. The VQA task is required to provide the correct answer to a given problem with a corresponding image. The aim of this paper is to design a complete framework and algorithm for image fragmentation information question answering under the basic background of visual question answering task. The main research includes image feature extraction, question text feature extraction, multi-modal feature fusion and answer reasoning. Deep neural network is constructed to extract features for representing images and problem information. Attention mechanism and variational inference method are combined to fusion two modal features of image and problem and reason answers. Experiment results show that the model can effectively extract and understand multi-modal fragmentation information, and improve the accuracy of VQA.
2018, 55(12): 2611-2619.
DOI: 10.7544/issn1000-1239.2018.20180575
Abstract：
The accuracy and stability of clustering will be obviously improved when a lot of independent clustering results for the same data set are aggregated by utilizing the principle of wisdom of crowds. In this paper, clustering ensemble algorithm with cluster connection based on wisdom of crowds (CECWOC) is proposed. Firstly, the independent clustering results are produced by the different clustering algorithms, which is guided by utilizing the independency, decentralization, diversity of wisdom of crowds. Secondly, the clustering ensemble algorithm based on connecting triple is developed to grouping aggregate the produced independent clusters, and the obtained results are aggregated again and the final cluster set is produced. The advantages of proposed algorithm are that: 1)The produced clusters by base clustering is grouping aggregated and weights of clusters are adjusted so that the selection of clusters is avoided, as a result, information on the produced clusters are not ignored; 2)Similarities of data are computed by using connected triple algorithm, the relations of data that their similarities are zero can be used. The experimental results at the different data sets show that the proposed algorithm can obtain the more accurate and stable results than other clustering ensemble algorithms, including the ones based on framework of wisdom of crowds.
The accuracy and stability of clustering will be obviously improved when a lot of independent clustering results for the same data set are aggregated by utilizing the principle of wisdom of crowds. In this paper, clustering ensemble algorithm with cluster connection based on wisdom of crowds (CECWOC) is proposed. Firstly, the independent clustering results are produced by the different clustering algorithms, which is guided by utilizing the independency, decentralization, diversity of wisdom of crowds. Secondly, the clustering ensemble algorithm based on connecting triple is developed to grouping aggregate the produced independent clusters, and the obtained results are aggregated again and the final cluster set is produced. The advantages of proposed algorithm are that: 1)The produced clusters by base clustering is grouping aggregated and weights of clusters are adjusted so that the selection of clusters is avoided, as a result, information on the produced clusters are not ignored; 2)Similarities of data are computed by using connected triple algorithm, the relations of data that their similarities are zero can be used. The experimental results at the different data sets show that the proposed algorithm can obtain the more accurate and stable results than other clustering ensemble algorithms, including the ones based on framework of wisdom of crowds.
2018, 55(12): 2620-2636.
DOI: 10.7544/issn1000-1239.2018.20160940
Abstract：
In view of the problem that the existing attack scenario construction methods are not accurate due to the lack of consideration of alarm missing and alarm redundancy, a new attack scenario construction method based on causal knowledge net is put forward. The causal knowledge net is composed of causal relationship and causal knowledge. Firstly, the causal relationship of single-step attacks is defined according to the expert knowledge, and then the real alarms are utilized to mine the causal knowledge, which can be used to quantitatively describe the causal relationship. In particular, the significance testing mean is designed to guarantee the consistency and accuracy of the causal relationship as well as causal knowledge among the mining causal knowledge. Additionally, the attack scenario construction method can be divided into two different steps with the help of causal knowledge net: the initiatory attack scenario can be obtained by means of alarm mapping and clustering in the first step, and in the second step, the initiatory attack scenario is reconstructed and the intact attack scenario is achieved by taking advantage of the theory named maximum a posteriori estimation. Experimental results show that the proposed method can improve the accuracy of attack scenario construction by combining the advantages of expert knowledge and data mining.
In view of the problem that the existing attack scenario construction methods are not accurate due to the lack of consideration of alarm missing and alarm redundancy, a new attack scenario construction method based on causal knowledge net is put forward. The causal knowledge net is composed of causal relationship and causal knowledge. Firstly, the causal relationship of single-step attacks is defined according to the expert knowledge, and then the real alarms are utilized to mine the causal knowledge, which can be used to quantitatively describe the causal relationship. In particular, the significance testing mean is designed to guarantee the consistency and accuracy of the causal relationship as well as causal knowledge among the mining causal knowledge. Additionally, the attack scenario construction method can be divided into two different steps with the help of causal knowledge net: the initiatory attack scenario can be obtained by means of alarm mapping and clustering in the first step, and in the second step, the initiatory attack scenario is reconstructed and the intact attack scenario is achieved by taking advantage of the theory named maximum a posteriori estimation. Experimental results show that the proposed method can improve the accuracy of attack scenario construction by combining the advantages of expert knowledge and data mining.
2018, 55(12): 2637-2650.
DOI: 10.7544/issn1000-1239.2018.20170773
Abstract：
E-government is evolving along with the development of national informatization. The e-government network security has become an important research field of national information security. However, the traditional LAN or Internet model cannot meet the need of e-government network based on region and domain management in China. The IATF(information assurance technical framework)theory originating from the US National Security Agency has become a reference for the design of network security architectures in many countries. However, the IATF model still cannot be well applied to the features and security management requirements of e-government network of China. At present, China’s e-government network has its own features characterized with the hierarchical domain protection architecture, classified controlled access requirements and graded responsibility management approach. Based on the in-depth analysis of e-government network status and requirements, a trust interconnection control (TIC) model for e-government network security is proposed to improve the trust control system. In the TIC model, three architecture designs of e-government network security are introduced, including the peer interconnectionmodel, hierarchical interconnection model and hybrid interconnection model, and the key technologies are designed in detail, such as cross-domain trusted transfer, inter-domain security supervision and whole-process strategy control etc. Finally, the TIC model is evaluated by analytic hierarchy process (AHP) method. The evaluation results show that in the complex e-government network, the TIC model can be suitable for the architecture design of the e-government network security. The key technologies in TIC can provide a valuable reference for the construction of the security system of the e-government networks and the implementation of the relevant products.
E-government is evolving along with the development of national informatization. The e-government network security has become an important research field of national information security. However, the traditional LAN or Internet model cannot meet the need of e-government network based on region and domain management in China. The IATF(information assurance technical framework)theory originating from the US National Security Agency has become a reference for the design of network security architectures in many countries. However, the IATF model still cannot be well applied to the features and security management requirements of e-government network of China. At present, China’s e-government network has its own features characterized with the hierarchical domain protection architecture, classified controlled access requirements and graded responsibility management approach. Based on the in-depth analysis of e-government network status and requirements, a trust interconnection control (TIC) model for e-government network security is proposed to improve the trust control system. In the TIC model, three architecture designs of e-government network security are introduced, including the peer interconnectionmodel, hierarchical interconnection model and hybrid interconnection model, and the key technologies are designed in detail, such as cross-domain trusted transfer, inter-domain security supervision and whole-process strategy control etc. Finally, the TIC model is evaluated by analytic hierarchy process (AHP) method. The evaluation results show that in the complex e-government network, the TIC model can be suitable for the architecture design of the e-government network security. The key technologies in TIC can provide a valuable reference for the construction of the security system of the e-government networks and the implementation of the relevant products.
2018, 55(12): 2651-2663.
DOI: 10.7544/issn1000-1239.2018.20170651
Abstract：
Wireless sensor networks have some obvious characteristics, such as communication range is limited, energy-constraint, network is vulnerable et al. Group key agreement in this environment requires a cross-cluster, and computation and communication overhead are lightweight and highly safe group key agreement protocol. Aiming at these demands, the paper proposes a cross-domain lightweight asymmetric group key agreement, in order to establish a safe and efficient group communication channel among sensor nodes. Firstly, the protocol establishes the secret information among the cluster heads, and the cluster head as the bridge node to realize the sensor nodes in different cluster have the same group key information, thus realizing the cross cluster asymmetric group key agreement. The whole network node can share the secret information with the internal nodes of the group, which realizes the group security communication mechanism of the message sender unconstraint; proposed an asymmetric calculation to achieve computation and communication migration technologies to ensure that the sensor nodes are lightweight computing and communication consumption. For our asymmetric GKA protocol, the key confirmation is simple and requires no additional rounds if the protocol has been correctly executed. Proven and analysis show that the proposed protocol has the advantages in security and energy consumption.
Wireless sensor networks have some obvious characteristics, such as communication range is limited, energy-constraint, network is vulnerable et al. Group key agreement in this environment requires a cross-cluster, and computation and communication overhead are lightweight and highly safe group key agreement protocol. Aiming at these demands, the paper proposes a cross-domain lightweight asymmetric group key agreement, in order to establish a safe and efficient group communication channel among sensor nodes. Firstly, the protocol establishes the secret information among the cluster heads, and the cluster head as the bridge node to realize the sensor nodes in different cluster have the same group key information, thus realizing the cross cluster asymmetric group key agreement. The whole network node can share the secret information with the internal nodes of the group, which realizes the group security communication mechanism of the message sender unconstraint; proposed an asymmetric calculation to achieve computation and communication migration technologies to ensure that the sensor nodes are lightweight computing and communication consumption. For our asymmetric GKA protocol, the key confirmation is simple and requires no additional rounds if the protocol has been correctly executed. Proven and analysis show that the proposed protocol has the advantages in security and energy consumption.
2018, 55(12): 2664-2673.
DOI: 10.7544/issn1000-1239.2018.20170757
Abstract：
Steganography is the art of covert communication technique by regular-looking media such as images, videos and texts, and the capability of anti-steganalysis is an important factor to measure its security. However, the research of steganography security develops slowly after the introduction of STC (syndrome tellis codes). Most of the existing studies are the completions and supplements of distortion function. In fact, there are two main factors that affect steganography security significantly: one is the embedding-related factor, such as payload or embedding algorithm; the other is the occultation effect of carrier, namely carrier security. In this paper, we study the carrier security by analyzing the relationship between steganography security and the carrier feature based on co-occurrence probability of image residual. We present an evaluation method of security for image carrier based on carrier distance of clustering center. The experimental results show that, by the security evaluation method we propose, the anti-steganalysis capacity of steganography is improved effectively by carrier selection. By the testing of the images under different databases, steganography algorithms, payloads and steganalysis features, the steganography based on our method enjoys a higher carrier security than random carrier selection and the average error rate of detection is improved about 3.8 to 11.8 percentage points.
Steganography is the art of covert communication technique by regular-looking media such as images, videos and texts, and the capability of anti-steganalysis is an important factor to measure its security. However, the research of steganography security develops slowly after the introduction of STC (syndrome tellis codes). Most of the existing studies are the completions and supplements of distortion function. In fact, there are two main factors that affect steganography security significantly: one is the embedding-related factor, such as payload or embedding algorithm; the other is the occultation effect of carrier, namely carrier security. In this paper, we study the carrier security by analyzing the relationship between steganography security and the carrier feature based on co-occurrence probability of image residual. We present an evaluation method of security for image carrier based on carrier distance of clustering center. The experimental results show that, by the security evaluation method we propose, the anti-steganalysis capacity of steganography is improved effectively by carrier selection. By the testing of the images under different databases, steganography algorithms, payloads and steganalysis features, the steganography based on our method enjoys a higher carrier security than random carrier selection and the average error rate of detection is improved about 3.8 to 11.8 percentage points.
2018, 55(12): 2674-2684.
DOI: 10.7544/issn1000-1239.2018.20170787
Abstract：
As more and more items are tagged with RFID tags. The grouping-proof technology which is used to produce a coexistence evidence with a group of related items is becoming more and more widely used. In the RFID grouping-proof protocol, how to generate reliable grouping-proof without threat to the tag information security and how to improve the protocol efficiency are search hotspots in RFID security area. In the off-line grouping-proof protocol, the proof data generally verified by the verifier to guarantee the privacy and security of tag information, the reader is only used to collect the grouping proof data, which cut down the responding speed to the illegal proof data in protocol. To enhance the grouping-proof efficiency and prevent DoP (deny of proof) attack, a RFID anonymous grouping-proof protocol using dual-layer verification (AGPDL) has been proposed. The AGPDL uses the elliptic curve cryptosystem as an encryption means. In this protocol, dual-layer verification is used. Reader are authorized to verify the validity of group-proof without knowing the identity of tags. After that, the final verification process and tag authentication are finished by the verifier in background server. Through the security and performance analysis, the AGPDL can provide the security and privacy about tag’s information, prevent the replay and impersonate attack, Furthermore, it has the ability to prevent the system overhead caused by invalid submission of grouping-proofs from reader with better scalability.
As more and more items are tagged with RFID tags. The grouping-proof technology which is used to produce a coexistence evidence with a group of related items is becoming more and more widely used. In the RFID grouping-proof protocol, how to generate reliable grouping-proof without threat to the tag information security and how to improve the protocol efficiency are search hotspots in RFID security area. In the off-line grouping-proof protocol, the proof data generally verified by the verifier to guarantee the privacy and security of tag information, the reader is only used to collect the grouping proof data, which cut down the responding speed to the illegal proof data in protocol. To enhance the grouping-proof efficiency and prevent DoP (deny of proof) attack, a RFID anonymous grouping-proof protocol using dual-layer verification (AGPDL) has been proposed. The AGPDL uses the elliptic curve cryptosystem as an encryption means. In this protocol, dual-layer verification is used. Reader are authorized to verify the validity of group-proof without knowing the identity of tags. After that, the final verification process and tag authentication are finished by the verifier in background server. Through the security and performance analysis, the AGPDL can provide the security and privacy about tag’s information, prevent the replay and impersonate attack, Furthermore, it has the ability to prevent the system overhead caused by invalid submission of grouping-proofs from reader with better scalability.
2018, 55(12): 2685-2702.
DOI: 10.7544/issn1000-1239.2018.20170587
Abstract：
Though electronic commerce companies adopt multifarious reputation evaluation mechanisms to guarantee trust between customers and sellers (or customers and platforms), these reputation evaluation systems are still frequently attacked. These attacks have led the reputation ranking and recommendation rankings of sellers to be manipulated. Therefore, large numbers of honest consumers are misled to purchase low quality products. It has been mentioned that overall consideration of trust and distrust information can improve customers’ ability in defensing against reputation attacks. However, existing works have limitations such as “the trust information and distrust information are less fused”, “one advisor list is used in evaluating all sellers”, which leads to the lack of pertinence and inaccuracy of sellers’ reputation evaluation. We propose a new defensing strategy called T&D. This strategy considers the trustworthy facet as well as the untrustworthy facet of customers. In addition, this strategy offers a whitelist (which stores several most trustworthy reviewers) and a blacklist (which stores several most untrustworthy reviewers) for customers. Based on the whitelist that is purified using the blacklist, honest customers can find the most trustworthy buyers and evaluate the candidate sellers according to its own experience and ratings of these trustworthy reviewers. Simulated experimental results show that our proposed strategy significantly outperforms state-of-the-art baselines in evaluation accuracy and stability.
Though electronic commerce companies adopt multifarious reputation evaluation mechanisms to guarantee trust between customers and sellers (or customers and platforms), these reputation evaluation systems are still frequently attacked. These attacks have led the reputation ranking and recommendation rankings of sellers to be manipulated. Therefore, large numbers of honest consumers are misled to purchase low quality products. It has been mentioned that overall consideration of trust and distrust information can improve customers’ ability in defensing against reputation attacks. However, existing works have limitations such as “the trust information and distrust information are less fused”, “one advisor list is used in evaluating all sellers”, which leads to the lack of pertinence and inaccuracy of sellers’ reputation evaluation. We propose a new defensing strategy called T&D. This strategy considers the trustworthy facet as well as the untrustworthy facet of customers. In addition, this strategy offers a whitelist (which stores several most trustworthy reviewers) and a blacklist (which stores several most untrustworthy reviewers) for customers. Based on the whitelist that is purified using the blacklist, honest customers can find the most trustworthy buyers and evaluate the candidate sellers according to its own experience and ratings of these trustworthy reviewers. Simulated experimental results show that our proposed strategy significantly outperforms state-of-the-art baselines in evaluation accuracy and stability.
2018, 55(12): 2703-2714.
DOI: 10.7544/issn1000-1239.2018.20170579
Abstract：
Video feature learning by deep neural network has become a hot research topic in video semantic analysis such as video object detection, motion recognition and video event detection. The topographic information of the video image plays an important role in describing the relationship between image and content. At the same time, it is helpful to improve the discriminability of the video feature expression by considering the characteristics of the video sequence with optimization. In this paper, an approach based on pre-training convolutional neural network with new topographic sparse encoder is proposed for video feature learning. This method has two stages: semi-supervised video image feature learning and supervised video sequence features optimization learning. In the semi-supervised video image feature learning stage, a new topographic sparse encoder is presented and used to pre-train neural networks, so that the characteristic expression of the video image can reflect the topographic information of the image, and a logistic regression is used to fine-tune the networks parameters using video concept label for video image feature learning. In the supervised video sequence feature optimization learning stage, a fully connected layer for feature learning of video sequence is constructed in order to express the feature of video sequence reasonably. A logistic regression constraint is established to adjust the network parameters in order that the discriminative feature of video sequence can be obtained. The experiments for relative methods are carried out on typical video datasets. The results show that the proposed method has better discriminability for the expression of video features, and can improve the accuracy of video semantic concept detection effectively.
Video feature learning by deep neural network has become a hot research topic in video semantic analysis such as video object detection, motion recognition and video event detection. The topographic information of the video image plays an important role in describing the relationship between image and content. At the same time, it is helpful to improve the discriminability of the video feature expression by considering the characteristics of the video sequence with optimization. In this paper, an approach based on pre-training convolutional neural network with new topographic sparse encoder is proposed for video feature learning. This method has two stages: semi-supervised video image feature learning and supervised video sequence features optimization learning. In the semi-supervised video image feature learning stage, a new topographic sparse encoder is presented and used to pre-train neural networks, so that the characteristic expression of the video image can reflect the topographic information of the image, and a logistic regression is used to fine-tune the networks parameters using video concept label for video image feature learning. In the supervised video sequence feature optimization learning stage, a fully connected layer for feature learning of video sequence is constructed in order to express the feature of video sequence reasonably. A logistic regression constraint is established to adjust the network parameters in order that the discriminative feature of video sequence can be obtained. The experiments for relative methods are carried out on typical video datasets. The results show that the proposed method has better discriminability for the expression of video features, and can improve the accuracy of video semantic concept detection effectively.
2018, 55(12): 2715-2724.
DOI: 10.7544/issn1000-1239.2018.20170766
Abstract：
Bridge rules provide an important mechanism describing semantic mapping and propagating knowledge for D3L (distributed dynamic description logics). The current research focuses on the homogeneous bridge rules which only contain atomic elements. In this paper, the research is extended to the D3L reasoning problem with the heterogeneous bridge rules which contain composite elements in the contained end. The regularity of distributed knowledge base is defined. Through the alternation of the bridge rules and transforming different forms into existing language mechanism, we present a algorithm which can convert the D3L knowledge base with dynamic description logic DSROIQ as local ontology language into a single DSROIQ knowledge base. Then we study the properties of the algorithm. We prove that the algorithm will terminate in polynomial time and the satisfiability of the target knowledge base is equivalent to the satisfiability of the original knowledge base. Thus, we prove that the worst-case time complexity of the centralized reasoning on regular D3L knowledge base with such bridge rules is the same as that on single DSROIQ knowledge base. The method proposed in this paper makes the reasoning for D3L to obtain the same worst-case time complexity as the existing distributed reasoning methods and solves the problem that the latter can not handle heterogeneous composite bridge rules.
Bridge rules provide an important mechanism describing semantic mapping and propagating knowledge for D3L (distributed dynamic description logics). The current research focuses on the homogeneous bridge rules which only contain atomic elements. In this paper, the research is extended to the D3L reasoning problem with the heterogeneous bridge rules which contain composite elements in the contained end. The regularity of distributed knowledge base is defined. Through the alternation of the bridge rules and transforming different forms into existing language mechanism, we present a algorithm which can convert the D3L knowledge base with dynamic description logic DSROIQ as local ontology language into a single DSROIQ knowledge base. Then we study the properties of the algorithm. We prove that the algorithm will terminate in polynomial time and the satisfiability of the target knowledge base is equivalent to the satisfiability of the original knowledge base. Thus, we prove that the worst-case time complexity of the centralized reasoning on regular D3L knowledge base with such bridge rules is the same as that on single DSROIQ knowledge base. The method proposed in this paper makes the reasoning for D3L to obtain the same worst-case time complexity as the existing distributed reasoning methods and solves the problem that the latter can not handle heterogeneous composite bridge rules.
2018, 55(12): 2725-2733.
DOI: 10.7544/issn1000-1239.2018.20170357
Abstract：
The loss function of AUC optimization involves pair-wise instances coming from different classes, so the objective functions of AUC methods, depending on the sum of pair-wise losses, are quadratic in the number of training examples. As a result, the objective functions of this type can not be directly solved through conventional online learning methods. The existing online AUC maximization methods focus on avoiding the direct calculation of all pair-wise loss functions, in order to reduce the problem sizes and achieve the online AUC optimization. To further solve the AUC optimization problems described above, we propose a novel AUC objective function that is only linear in the number of training examples. Theoretical analysis shows the minimization of the proposed objective function is equivalent to that of the objective function for AUC optimization by the combination of L2 regularization and least square surrogate loss. Based on this new objective function, we obtain the method named linear online AUC maximization (LOAM). According to different updating strategies for classifiers, we develop two algorithms for LOAM method: LOAM\-{ILSC} and LOAM\-{Ada}. Experimental results show that, compared with the rival methods, LOAM\-{ILSC} can achieve better AUC performance, and LOAM\-{Ada} is more effective and efficient to handle real-time or high dimensional learning tasks.
The loss function of AUC optimization involves pair-wise instances coming from different classes, so the objective functions of AUC methods, depending on the sum of pair-wise losses, are quadratic in the number of training examples. As a result, the objective functions of this type can not be directly solved through conventional online learning methods. The existing online AUC maximization methods focus on avoiding the direct calculation of all pair-wise loss functions, in order to reduce the problem sizes and achieve the online AUC optimization. To further solve the AUC optimization problems described above, we propose a novel AUC objective function that is only linear in the number of training examples. Theoretical analysis shows the minimization of the proposed objective function is equivalent to that of the objective function for AUC optimization by the combination of L2 regularization and least square surrogate loss. Based on this new objective function, we obtain the method named linear online AUC maximization (LOAM). According to different updating strategies for classifiers, we develop two algorithms for LOAM method: LOAM\-{ILSC} and LOAM\-{Ada}. Experimental results show that, compared with the rival methods, LOAM\-{ILSC} can achieve better AUC performance, and LOAM\-{Ada} is more effective and efficient to handle real-time or high dimensional learning tasks.
2018, 55(12): 2734-2740.
DOI: 10.7544/issn1000-1239.2018.20170529
Abstract：
Constraint propagation is one of the key methods of constraint programming, and it can be used for solving constraint satisfaction problems and also deal with industrial modeling issues. In recent years, simple tabular reduction (STR) algorithms have been frequently used in some constraint propagation algorithms to cut down the consumption of constraint table space, and at the same time increase running speed of the generalised arc consistent (GAC) algorithm. For the past few years, short support method was the most frequently used as a table compression method in constraint propagation algorithms. This method can propagate more constraints than original STR algorithms especially when the memory is small. But when the compression ratio is low, short support method improving the running speed effect is not obvious. In this paper, we present a new algorithm to compress constraint table, called simple tabular reduction optimization (STRO), combined short support compression method and bit-wise operation. STRO algorithm improves the running speed of STR algorithm, and at the same time, the compression of space effect is better. Experimental results show that when the average size of the table is not particularly small, STRO algorithm is faster and more efficient than ShortSTR2 and STR2 algorithm; compared with STRbit algorithm, the compression rate of STRO algorithm is bigger, and it can save more space and replace STRbit algorithm on time.
Constraint propagation is one of the key methods of constraint programming, and it can be used for solving constraint satisfaction problems and also deal with industrial modeling issues. In recent years, simple tabular reduction (STR) algorithms have been frequently used in some constraint propagation algorithms to cut down the consumption of constraint table space, and at the same time increase running speed of the generalised arc consistent (GAC) algorithm. For the past few years, short support method was the most frequently used as a table compression method in constraint propagation algorithms. This method can propagate more constraints than original STR algorithms especially when the memory is small. But when the compression ratio is low, short support method improving the running speed effect is not obvious. In this paper, we present a new algorithm to compress constraint table, called simple tabular reduction optimization (STRO), combined short support compression method and bit-wise operation. STRO algorithm improves the running speed of STR algorithm, and at the same time, the compression of space effect is better. Experimental results show that when the average size of the table is not particularly small, STRO algorithm is faster and more efficient than ShortSTR2 and STR2 algorithm; compared with STRbit algorithm, the compression rate of STRO algorithm is bigger, and it can save more space and replace STRbit algorithm on time.
2018, 55(12): 2741-2752.
DOI: 10.7544/issn1000-1239.2018.20170336
Abstract：
To resolve the problem that the existing single image-based noise level estimation (SNLE) algorithms are of poor anti-interference ability and low execution efficiency, a multi-image based noise level estimation (MNLE) algorithm using prior knowledge of similar images is proposed. Specifically, a set of distorted images is first constructed by adding different pre-defined noise levels to some representative noise-free images, and several natural statistical values extracted from each distorted image are used to form the noise level-aware feature vector. Then, the feature vector extracted from each distorted image and its corresponding noise level are used to construct a sample database. During the noise level estimation stage, the feature vector of an image to be estimated is extracted by the same method as the preparation stage, and several feature vectors similar to the extracted feature vector and their corresponding noise levels are chosen from the sample database. As such, the noise level of the image to be estimated is estimated with a weighted average approach. Extensive experimental results show that the proposed MNLE algorithm not only is of high efficiency but also has stable prediction accuracy at high, medium, and low noise levels. For the Gaussian noise mixed with impulse noise or Poisson noise, it also has good anti-interference ability.
To resolve the problem that the existing single image-based noise level estimation (SNLE) algorithms are of poor anti-interference ability and low execution efficiency, a multi-image based noise level estimation (MNLE) algorithm using prior knowledge of similar images is proposed. Specifically, a set of distorted images is first constructed by adding different pre-defined noise levels to some representative noise-free images, and several natural statistical values extracted from each distorted image are used to form the noise level-aware feature vector. Then, the feature vector extracted from each distorted image and its corresponding noise level are used to construct a sample database. During the noise level estimation stage, the feature vector of an image to be estimated is extracted by the same method as the preparation stage, and several feature vectors similar to the extracted feature vector and their corresponding noise levels are chosen from the sample database. As such, the noise level of the image to be estimated is estimated with a weighted average approach. Extensive experimental results show that the proposed MNLE algorithm not only is of high efficiency but also has stable prediction accuracy at high, medium, and low noise levels. For the Gaussian noise mixed with impulse noise or Poisson noise, it also has good anti-interference ability.
2018, 55(12): 2753-2763.
DOI: 10.7544/issn1000-1239.2018.20170523
Abstract：
Active Demons algorithm based on fractional differential has been proved to be effective for non-rigid image registration, and can solve the problem of low accuracy and low efficiency in image registration for intensity uniformity or weak texture region. However, the optimal order of fractional differential operator in the algorithm needs to be selected manually by multiple experiments, lack of order adaptive in image registration. Aiming at the problem, this paper proposes a new Active Demons algorithm based on multi-resolution and adaptive fractional differential. Firstly, an adaptive fractional order mathematical model is constructed based on the gradient magnitude and information entropy of image, therefore the optimal order and differential dynamic template are adjusted adaptively. Additionally, multi-resolution strategy is introduced to adaptive fractional differential Active Demons algorithm, therefore the registration efficiency is improved. Theory analysis and experimental results show that the proposed algorithm is capable of registrating images with intensity uniformity, weak edge and weak texture. And the optimal order of fractional differential operator can be calculated adaptively. Furthermore, the presented method can avoid falling into local optimum, thus the accuracy and efficiency of registration can be improved.
Active Demons algorithm based on fractional differential has been proved to be effective for non-rigid image registration, and can solve the problem of low accuracy and low efficiency in image registration for intensity uniformity or weak texture region. However, the optimal order of fractional differential operator in the algorithm needs to be selected manually by multiple experiments, lack of order adaptive in image registration. Aiming at the problem, this paper proposes a new Active Demons algorithm based on multi-resolution and adaptive fractional differential. Firstly, an adaptive fractional order mathematical model is constructed based on the gradient magnitude and information entropy of image, therefore the optimal order and differential dynamic template are adjusted adaptively. Additionally, multi-resolution strategy is introduced to adaptive fractional differential Active Demons algorithm, therefore the registration efficiency is improved. Theory analysis and experimental results show that the proposed algorithm is capable of registrating images with intensity uniformity, weak edge and weak texture. And the optimal order of fractional differential operator can be calculated adaptively. Furthermore, the presented method can avoid falling into local optimum, thus the accuracy and efficiency of registration can be improved.
2018, 55(12): 2764-2774.
DOI: 10.7544/issn1000-1239.2018.20170891
Abstract：
Gesture interaction based on data glove is an important way to realize human computer interaction . The effect of gesture recognition and spatial positioning has an important influence on the immersion and experience of virtual environment interaction tasks. The complex structure of typical products lead to high complexity of modeling and gesture recognition algorithm，and it’s hard to balance the cost and efficiency.Aiming at the requirements of application,we come up with a design of data glove system which combines the sensor data with visual information in this paper. A rapid mapping that convert the finger’s curvature value to virtual hand joints position is realized by a hierarchical template matching method,which simplifies the process of gesture’s modeling and matching effectively. In order to obtain the spatial information for gesture interaction, an efficient feature-checking Hough algorithm is proposed to locate the data glove. We use the camera to capture the circular contour of the luminous ball on the glove and calculate it’s position in three-dimensional space, and use a modified Hough-Circle-Detection algorithm based on diameter equation to reduce the circular detection complexity. For continuous frame image, the credibility based feature-checking rule is adopted to improve the overall efficiency of the algorithm.This scheme was applied in immersive virtual reality experiment system for middle school. Experimental results show that the system is stable and has a good interactive experience.
Gesture interaction based on data glove is an important way to realize human computer interaction . The effect of gesture recognition and spatial positioning has an important influence on the immersion and experience of virtual environment interaction tasks. The complex structure of typical products lead to high complexity of modeling and gesture recognition algorithm，and it’s hard to balance the cost and efficiency.Aiming at the requirements of application,we come up with a design of data glove system which combines the sensor data with visual information in this paper. A rapid mapping that convert the finger’s curvature value to virtual hand joints position is realized by a hierarchical template matching method,which simplifies the process of gesture’s modeling and matching effectively. In order to obtain the spatial information for gesture interaction, an efficient feature-checking Hough algorithm is proposed to locate the data glove. We use the camera to capture the circular contour of the luminous ball on the glove and calculate it’s position in three-dimensional space, and use a modified Hough-Circle-Detection algorithm based on diameter equation to reduce the circular detection complexity. For continuous frame image, the credibility based feature-checking rule is adopted to improve the overall efficiency of the algorithm.This scheme was applied in immersive virtual reality experiment system for middle school. Experimental results show that the system is stable and has a good interactive experience.
2018, 55(12): 2775-2784.
DOI: 10.7544/issn1000-1239.2018.20170581
Abstract：
Image representations derived from pre-trained convolutional neural networks (CNNs) have become the new state of the art in the task of image retrieval. But these methods are all based on image global representations and can’t be applied to the retrieval of query objects which only occupy the part area of the retrieved images. To solve these problems, this work explores the suitability for object retrieval of small query objects which only occupy part area of the retrieved images using pre-trained fully convolutional networks. First, we take advantage of the fully convolutional networks without the restriction of the size of input image，and given retrieved images，feature matrix representations are derived by fully convolutional networks. Second, given the query object, the feature can also be derived by the fully convolutional networks. Finally, the feature of query object is matched with each feature of the feature matrix of the retrieved image, and the similarity and optimal matching location are obtained. We further investigate the suitability of the multi-scale, multi-ratio transformation for different sizes of query object in the retrieved image. Experimental results on the benchmark dataset Oxford5K show that our method outperforms other state-of-the-art methods. We further demonstrate its scalability and efficacy on the Logo dataset which is collected randomly from the Internet.
Image representations derived from pre-trained convolutional neural networks (CNNs) have become the new state of the art in the task of image retrieval. But these methods are all based on image global representations and can’t be applied to the retrieval of query objects which only occupy the part area of the retrieved images. To solve these problems, this work explores the suitability for object retrieval of small query objects which only occupy part area of the retrieved images using pre-trained fully convolutional networks. First, we take advantage of the fully convolutional networks without the restriction of the size of input image，and given retrieved images，feature matrix representations are derived by fully convolutional networks. Second, given the query object, the feature can also be derived by the fully convolutional networks. Finally, the feature of query object is matched with each feature of the feature matrix of the retrieved image, and the similarity and optimal matching location are obtained. We further investigate the suitability of the multi-scale, multi-ratio transformation for different sizes of query object in the retrieved image. Experimental results on the benchmark dataset Oxford5K show that our method outperforms other state-of-the-art methods. We further demonstrate its scalability and efficacy on the Logo dataset which is collected randomly from the Internet.
2018, 55(12): 2785-2793.
DOI: 10.7544/issn1000-1239.2018.20170327
Abstract：
Deep networks have been successfully applied to visual tracking by learning a generic representation offline from numerous training images. However, the features of the convolutional neural network abstraction algorithm are lack of spatio-temporal context information and the offline training is time-consuming. To tackle the above issues, an online convolution network tracking via spatio-temporal context is proposed, adopting the spatio-temporal context as the every order filter in convolutional neural network. Firstly, the initial target is normalized and the target confidence map is extracted. In the process of tracking, the spatio-temporal information is updated to obtain the spatio-temporal context model. The first layer utilizes the updated model to convolve the input and performs sliding window on the convolution result. The second layer convolves the fetch results by spatio-temporal model respectively, extracts the simple target abstract features, and then the convolution result of the simple layer is superposed to the deep level target expression. Finally, the target tracking is realized by the particle filter tracking framework. Our convolutional networks have a lightweight structure and perform favorably against several state-of-the-art methods on OTB-2013 and OTB-2015. As documented in the experimental results, the deep abstract feature extracted by online convolution network structure combining with spatio-temporal context model, can preserve related spatio-temporal information and then the tracking efficiency under complex background is improved.
Deep networks have been successfully applied to visual tracking by learning a generic representation offline from numerous training images. However, the features of the convolutional neural network abstraction algorithm are lack of spatio-temporal context information and the offline training is time-consuming. To tackle the above issues, an online convolution network tracking via spatio-temporal context is proposed, adopting the spatio-temporal context as the every order filter in convolutional neural network. Firstly, the initial target is normalized and the target confidence map is extracted. In the process of tracking, the spatio-temporal information is updated to obtain the spatio-temporal context model. The first layer utilizes the updated model to convolve the input and performs sliding window on the convolution result. The second layer convolves the fetch results by spatio-temporal model respectively, extracts the simple target abstract features, and then the convolution result of the simple layer is superposed to the deep level target expression. Finally, the target tracking is realized by the particle filter tracking framework. Our convolutional networks have a lightweight structure and perform favorably against several state-of-the-art methods on OTB-2013 and OTB-2015. As documented in the experimental results, the deep abstract feature extracted by online convolution network structure combining with spatio-temporal context model, can preserve related spatio-temporal information and then the tracking efficiency under complex background is improved.
2018, 55(12): 2794-2809.
DOI: 10.7544/issn1000-1239.2018.20170756
Abstract：
The problem of differentially private data publishing has attracted considerable research attention in recent years. The current existing solutions, however, cannot effectively handle the release of high-dimensional data. That is because these methods suffer from curse of dimensionality and various domain sizes, which will lead to the lower utility of publication. To address the problems, this paper presents PrivHD (differentially private high dimensional data release) with junction tree, a differentially private method for publishing high-dimensional data. PrivHD firstly generates a Markov network with exponential mechanism, which employs the high-pass filter technique to reduce the candidate space in the sampling process. After that, based on the network, PrivHD obtains a complete cluster graph in terms of full triangulation and node elimination, and then relies on the cluster graph and maximum spanning tree method to construct a differentially private junction tree. Finally, PrivHD uses the post-processing technique to boost the noisy counts of marginal tables in each cluster in junction tree, and based on the boosted result, PrivHD produces the high-dimensional synthetic dataset. PrivHD is compared with the existing approaches such as PrivBayes, JTree on the different real datasets. The experimental results show that PrivHD is better than its competitors on k-way query and SVM classification.
The problem of differentially private data publishing has attracted considerable research attention in recent years. The current existing solutions, however, cannot effectively handle the release of high-dimensional data. That is because these methods suffer from curse of dimensionality and various domain sizes, which will lead to the lower utility of publication. To address the problems, this paper presents PrivHD (differentially private high dimensional data release) with junction tree, a differentially private method for publishing high-dimensional data. PrivHD firstly generates a Markov network with exponential mechanism, which employs the high-pass filter technique to reduce the candidate space in the sampling process. After that, based on the network, PrivHD obtains a complete cluster graph in terms of full triangulation and node elimination, and then relies on the cluster graph and maximum spanning tree method to construct a differentially private junction tree. Finally, PrivHD uses the post-processing technique to boost the noisy counts of marginal tables in each cluster in junction tree, and based on the boosted result, PrivHD produces the high-dimensional synthetic dataset. PrivHD is compared with the existing approaches such as PrivBayes, JTree on the different real datasets. The experimental results show that PrivHD is better than its competitors on k-way query and SVM classification.
