Email alert
2022 Vol. 44, No. 1
Display Method:
2022, 44(1): 168-177.
doi: 10.11999/JEIT200942
Abstract:
Due to the lack of unified human activity model and related specifications, the existing wearable human activity recognition technology uses different types, numbers and deployment locations of sensors, and affects its promotion and application. Based on the analysis of human activity skeleton characteristics and human activity mechanics, a human activity model based on Cartesian coordinates is established and the normalization method of activity sensor deployment location and activity data in the model is standardized; Secondly, a sliding window technique is introduced to establish a mapping method to convert human activity data into RGB bitmap, and a Convolutional Neural Network is designed for Human Activity Recognition (HAR-CNN); Finally, a HAR-CNN instance is created and experimentally tested based on the public human activity dataset Opportunity. The experimental results show that HAR-CNN achieves the F1 values of 90% and 92% for periodic repetitive activity and discrete human activity recognition, respectively, while the algorithm has good operational efficiency.
Due to the lack of unified human activity model and related specifications, the existing wearable human activity recognition technology uses different types, numbers and deployment locations of sensors, and affects its promotion and application. Based on the analysis of human activity skeleton characteristics and human activity mechanics, a human activity model based on Cartesian coordinates is established and the normalization method of activity sensor deployment location and activity data in the model is standardized; Secondly, a sliding window technique is introduced to establish a mapping method to convert human activity data into RGB bitmap, and a Convolutional Neural Network is designed for Human Activity Recognition (HAR-CNN); Finally, a HAR-CNN instance is created and experimentally tested based on the public human activity dataset Opportunity. The experimental results show that HAR-CNN achieves the F1 values of 90% and 92% for periodic repetitive activity and discrete human activity recognition, respectively, while the algorithm has good operational efficiency.
2022, 44(1): 178-186.
doi: 10.11999/JEIT201046
Abstract:
In order to improve the effect of image super-resolution reconstruction, the attention mechanism is introduced into Multi-level Residual Attention Network (MRAN) as the improved reconstruction network of Cycle Generation Countermeasure Network (CycleGAN) in this paper. A super-resolution reconstruction model MRA-GAN based on CycleGAN is proposed. The designed reconstruction network in MRA-GAN model is responsible for mapping from Low Resolution (LR) image to High Resolution (HR) image and the designed degradation network is responsible for reconstructing HR image to LR image. The LR discriminator is used to identify the real LR image which is obtained through the degraded network. The HR discriminator is used to identify the real HR image which is reconstructed by the reconstructed network. Moreover, the original discriminator and loss function of CycleGAN is improved. Experimental results verify that MRA-GAN model can obtain better Peak Signal to Noise Ratio (PSNR) and Structural SIMilarity (SSIM) than the existing deep learning based super-resolution algorithms.
In order to improve the effect of image super-resolution reconstruction, the attention mechanism is introduced into Multi-level Residual Attention Network (MRAN) as the improved reconstruction network of Cycle Generation Countermeasure Network (CycleGAN) in this paper. A super-resolution reconstruction model MRA-GAN based on CycleGAN is proposed. The designed reconstruction network in MRA-GAN model is responsible for mapping from Low Resolution (LR) image to High Resolution (HR) image and the designed degradation network is responsible for reconstructing HR image to LR image. The LR discriminator is used to identify the real LR image which is obtained through the degraded network. The HR discriminator is used to identify the real HR image which is reconstructed by the reconstructed network. Moreover, the original discriminator and loss function of CycleGAN is improved. Experimental results verify that MRA-GAN model can obtain better Peak Signal to Noise Ratio (PSNR) and Structural SIMilarity (SSIM) than the existing deep learning based super-resolution algorithms.
2022, 44(1): 187-194.
doi: 10.11999/JEIT210219
Abstract:
At present, the emergency search Unmanned Aerial Vehicle (UAV) cluster has problems such as low search efficiency, low coverage integrity, and poor stability of multi-unit network. In this regard, a terminal-routing UAV area search task planning strategy based on Optimized Fuzzy C-Means Algorithm (O-FCMA) combined with Optimize-Hybrid Particle Swarm Optimization (O-HPSO) algorithm is proposed. Based on the scope of UAV monitoring area, by establishing the spatial model of the search area, this paper further uses O-FCMA to divide the area geometrically, and uses O-HPSO to realize the path planning in the divided area, so as to realize the planning of the overall task of multi-UAV cluster search. The simulation experiment is performed. The use of O-HPSO combined with O-FCMA for passive UAV area search task is compared with ACO or simulated annealing algorithm combined with K clustering algorithm or FCMA, under the condition of ensuring the full coverage of the search area, active search compared with non-active UAV, the UAV decision time is reduced by 7%～21%, 16%～31%, and the search efficiency is increased by 7%～13%, 3%～7%. The method reduces effectively the UAV cluster decision time and improves the search efficiency.
At present, the emergency search Unmanned Aerial Vehicle (UAV) cluster has problems such as low search efficiency, low coverage integrity, and poor stability of multi-unit network. In this regard, a terminal-routing UAV area search task planning strategy based on Optimized Fuzzy C-Means Algorithm (O-FCMA) combined with Optimize-Hybrid Particle Swarm Optimization (O-HPSO) algorithm is proposed. Based on the scope of UAV monitoring area, by establishing the spatial model of the search area, this paper further uses O-FCMA to divide the area geometrically, and uses O-HPSO to realize the path planning in the divided area, so as to realize the planning of the overall task of multi-UAV cluster search. The simulation experiment is performed. The use of O-HPSO combined with O-FCMA for passive UAV area search task is compared with ACO or simulated annealing algorithm combined with K clustering algorithm or FCMA, under the condition of ensuring the full coverage of the search area, active search compared with non-active UAV, the UAV decision time is reduced by 7%～21%, 16%～31%, and the search efficiency is increased by 7%～13%, 3%～7%. The method reduces effectively the UAV cluster decision time and improves the search efficiency.
2022, 44(1): 195-202.
doi: 10.11999/JEIT201006
Abstract:
In order to make the network capture more effective content distinguish pedestrians, this paper proposes a multi-branch network based on Stepped feature space segmentation and Local Branch Attention Network (SLANet) mechanism to pay attention to the salient information of local images. First of all, a stepped branch attention module is introduced into the network. This module blocks the feature map horizontally in a stepped manner, and branch attention is used to assign different weights to each branch. Secondly, a multi-scale adaptive attention module is introduced into the network, which processes local features and adapts the size of the receptive field to adapt to images of different scales. Meanwhile, channel attention and spatial attention are combined to screen out the important features of the image. In the design of network, the multi-granularity network is used to combine the global feature with the local feature. Finally, the method is validated on three widely used person re-identification data sets Market-1501, DukeMTMC-reID and CUHK03. Among them, mAP and Rank-1 on market-1501 data set reach 88.1% and 95.6% respectively. The experimental results show that the proposed network model can improve the accuracy of person re-identification.
In order to make the network capture more effective content distinguish pedestrians, this paper proposes a multi-branch network based on Stepped feature space segmentation and Local Branch Attention Network (SLANet) mechanism to pay attention to the salient information of local images. First of all, a stepped branch attention module is introduced into the network. This module blocks the feature map horizontally in a stepped manner, and branch attention is used to assign different weights to each branch. Secondly, a multi-scale adaptive attention module is introduced into the network, which processes local features and adapts the size of the receptive field to adapt to images of different scales. Meanwhile, channel attention and spatial attention are combined to screen out the important features of the image. In the design of network, the multi-granularity network is used to combine the global feature with the local feature. Finally, the method is validated on three widely used person re-identification data sets Market-1501, DukeMTMC-reID and CUHK03. Among them, mAP and Rank-1 on market-1501 data set reach 88.1% and 95.6% respectively. The experimental results show that the proposed network model can improve the accuracy of person re-identification.
2022, 44(1): 203-210.
doi: 10.11999/JEIT200934
Abstract:
Point-Of-Interest (POI) recommendation in location-based social networks is an important way for people to find interesting locations. However, in reality, both the various user preference of locations in different regions and the high-dimensional historical check-in information make accurate and personalized POI recommendations extremely challenging. In this regard, a new type of recommendation algorithm for point-of-interest Partition Recommendation based on a category transfer Weighted Tensor Decomposition (WTD-PR) model is proposed. The proposed algorithm makes full use of the user’s historical visit information by combining the user’s continuous behavior and time characteristics to obtain the category transfer weight factor; Then, by improving the user-time-category tensor model and adding the category transfer weight to the tensor to predict the user’s preference category; Finally, the local and remote locations are divided according to the user’s historical access area, and the recommended areas are found based on the user’s current location. After that, location and social factors are introduced and combined with the candidate categories to make the recommendation of points of interest. Through comparative experiments on real data sets, the proposed algorithm is proved not only to be universal, but also superior to other comparison algorithms in terms of recommendation performance.
Point-Of-Interest (POI) recommendation in location-based social networks is an important way for people to find interesting locations. However, in reality, both the various user preference of locations in different regions and the high-dimensional historical check-in information make accurate and personalized POI recommendations extremely challenging. In this regard, a new type of recommendation algorithm for point-of-interest Partition Recommendation based on a category transfer Weighted Tensor Decomposition (WTD-PR) model is proposed. The proposed algorithm makes full use of the user’s historical visit information by combining the user’s continuous behavior and time characteristics to obtain the category transfer weight factor; Then, by improving the user-time-category tensor model and adding the category transfer weight to the tensor to predict the user’s preference category; Finally, the local and remote locations are divided according to the user’s historical access area, and the recommended areas are found based on the user’s current location. After that, location and social factors are introduced and combined with the candidate categories to make the recommendation of points of interest. Through comparative experiments on real data sets, the proposed algorithm is proved not only to be universal, but also superior to other comparison algorithms in terms of recommendation performance.
2022, 44(1): 211-220.
doi: 10.11999/JEIT201003
Abstract:
To relieve the problem of occlusion and misalignment caused by pose/view variations in real world, a new deep architecture named Diversified Local Attention Network (DLAN) for person Re-IDentification (Re-ID) is proposed in this paper. On the whole, a global branch and multiple local attention branches are designed following the backbone network, which simultaneously learn the pedestrians' global spatial structure and salient local features of different body parts. Furthermore, a novel Consistent Activation Penalty (CAP) is devised to constraint the output of local networks so as to obtain the complementary and diversified feature representations. Finally, the global and local features are fed into the classification network to form more comprehensive description of pedestrian via jointly learning. Utilizing Market1501, DukeMTMC-reID and CUHK03 datasets, the proposed DLAN model has reached 88.4%/95.1%, 79.5%/88.7% and 74.3%/76.3% (mAP/Rank-1) respectively, which are better than the compared methods. The experiments adequately verify the robustness and discriminability of the proposed method.
To relieve the problem of occlusion and misalignment caused by pose/view variations in real world, a new deep architecture named Diversified Local Attention Network (DLAN) for person Re-IDentification (Re-ID) is proposed in this paper. On the whole, a global branch and multiple local attention branches are designed following the backbone network, which simultaneously learn the pedestrians' global spatial structure and salient local features of different body parts. Furthermore, a novel Consistent Activation Penalty (CAP) is devised to constraint the output of local networks so as to obtain the complementary and diversified feature representations. Finally, the global and local features are fed into the classification network to form more comprehensive description of pedestrian via jointly learning. Utilizing Market1501, DukeMTMC-reID and CUHK03 datasets, the proposed DLAN model has reached 88.4%/95.1%, 79.5%/88.7% and 74.3%/76.3% (mAP/Rank-1) respectively, which are better than the compared methods. The experiments adequately verify the robustness and discriminability of the proposed method.
2022, 44(1): 221-229.
doi: 10.11999/JEIT200817
Abstract:
To solve the problems of lack of background knowledge and poor consistency of robot response in the existing human-computer interaction, a human-computer interaction model based on the ripple network of knowledge graph is proposed. In order to achieve a more natural and intelligent human-computer interaction system, this model simulates the real human-human interaction process. Firstly, the human-computer interaction affective friendliness is obtained by calculating the human-computer interaction emotional evaluation value and the human-computer interaction emotional certainty degree. Then, the external knowledge graph is introduced as the background knowledge of robots, and the dialogue entity is embedded into the ripple network of knowledge graph to obtain the potential interested entity content of the participants. Finally, considering the emotional friendliness and content friendliness, the robot response is given. The experimental results show that, compared with other models, robots that have background knowledge and consider emotional friendliness improve emotionality and coherence when interacting with human.
To solve the problems of lack of background knowledge and poor consistency of robot response in the existing human-computer interaction, a human-computer interaction model based on the ripple network of knowledge graph is proposed. In order to achieve a more natural and intelligent human-computer interaction system, this model simulates the real human-human interaction process. Firstly, the human-computer interaction affective friendliness is obtained by calculating the human-computer interaction emotional evaluation value and the human-computer interaction emotional certainty degree. Then, the external knowledge graph is introduced as the background knowledge of robots, and the dialogue entity is embedded into the ripple network of knowledge graph to obtain the potential interested entity content of the participants. Finally, considering the emotional friendliness and content friendliness, the robot response is given. The experimental results show that, compared with other models, robots that have background knowledge and consider emotional friendliness improve emotionality and coherence when interacting with human.
2022, 44(1): 230-236.
doi: 10.11999/JEIT200933
Abstract:
Traditional single-channel blind deconvolution method has the limitation that it can only separate two sources from a mixture. Considering this problem, a Single-Channel Blind Deconvolution algorithm based on optimized deep Convolutional generative adversarial networks (SCBDC) is proposed to separate and deconvolve more than three independent sources and mixing matrix only from a mixture. The experiments are carried on the occlusion Chinese character image datasets, four sources are randomly selected to be mixed with mixing matrix. Peak Signal to Noise Ratio (PSNR) and signal correlation index are combined to evaluate the separation effect. The result shows that the multiple sources can be effectively separated and deconvolved.
Traditional single-channel blind deconvolution method has the limitation that it can only separate two sources from a mixture. Considering this problem, a Single-Channel Blind Deconvolution algorithm based on optimized deep Convolutional generative adversarial networks (SCBDC) is proposed to separate and deconvolve more than three independent sources and mixing matrix only from a mixture. The experiments are carried on the occlusion Chinese character image datasets, four sources are randomly selected to be mixed with mixing matrix. Peak Signal to Noise Ratio (PSNR) and signal correlation index are combined to evaluate the separation effect. The result shows that the multiple sources can be effectively separated and deconvolved.
2022, 44(1): 237-244.
doi: 10.11999/JEIT200792
Abstract:
In order to make the fused multispectral images preserve the spectral information of the original Low-Resolution Multi-Spectral (LRMS) images as much as possible, and improve the spatial resolution effectively, a new pan-sharpening method based on multi-stream architecture and multi-scale is proposed. Firstly, This paper inputs the original MS image into the spectral feature extraction subnet to obtain its spectral features, and extracts the multi-directional gradient information and spatial structure information from the panchromatic images by the gradient operator and the convolution kernel. Then the extracted feature is added into the pyramid module with multi-stream fusion architecture for image reconstruction. The pyramid module is composed of multiple backbone networks, which can perform feature extraction under different spatial receptive fields, and can learn image information at multiple scales. Finally, a spatial spectrum prediction subnet is constructed to fuse the high-level features output by the pyramid module and the low-level features of the network front end to obtain multispectral images with high spatial resolution. Experiments on images obtained by WorldView-3 satellites show that the fusion images generated by the proposed method are superior to the most of advanced remote sensing image pan-sharpening methods in both subjective visual and objective evaluation indicators.
In order to make the fused multispectral images preserve the spectral information of the original Low-Resolution Multi-Spectral (LRMS) images as much as possible, and improve the spatial resolution effectively, a new pan-sharpening method based on multi-stream architecture and multi-scale is proposed. Firstly, This paper inputs the original MS image into the spectral feature extraction subnet to obtain its spectral features, and extracts the multi-directional gradient information and spatial structure information from the panchromatic images by the gradient operator and the convolution kernel. Then the extracted feature is added into the pyramid module with multi-stream fusion architecture for image reconstruction. The pyramid module is composed of multiple backbone networks, which can perform feature extraction under different spatial receptive fields, and can learn image information at multiple scales. Finally, a spatial spectrum prediction subnet is constructed to fuse the high-level features output by the pyramid module and the low-level features of the network front end to obtain multispectral images with high spatial resolution. Experiments on images obtained by WorldView-3 satellites show that the fusion images generated by the proposed method are superior to the most of advanced remote sensing image pan-sharpening methods in both subjective visual and objective evaluation indicators.
2022, 44(1): 245-253.
doi: 10.11999/JEIT200779
Abstract:
Most recommendation systems have a data sparsity problem, which limits the validity of the model that they use. However, the user’s comments on a commodity contain a lot of information. Emotional analysis of the comment text and the extraction of key factors for model learning can effectively alleviate the data sparsity problem, but only the use of comment data and ignore the main factors of the scoring data will affect the recommendation accuracy. To improve further the precision of recommendations, a deep model for the processing of Review Texts and Rating Matrices (RTRM) is proposed. The model extracts deep-level features and combines them to make rating predictions. Then, by using the pre-trained Electra model, the implicit expression of each comment is get, and combining with the deep emotion analysis and attention mechanism, the analysis of the comment text is realized from the context semantic level. It solves the problem that it is difficult to analyze the semantics of short text; User (item) reviews interact with a rating matrix to predict the user’s rating of a product in the fusion layer module. Finally, the Mean Square Error (MSE) is used to perform performance comparison experiments on 6 sets of data sets. Experimental results show that the performance of the proposed model outperforms significantly other systems on a variety of datasets, and the average prediction error is reduced by a maximum of 12.821%, the model is suitable for recommending accurate items to users.
Most recommendation systems have a data sparsity problem, which limits the validity of the model that they use. However, the user’s comments on a commodity contain a lot of information. Emotional analysis of the comment text and the extraction of key factors for model learning can effectively alleviate the data sparsity problem, but only the use of comment data and ignore the main factors of the scoring data will affect the recommendation accuracy. To improve further the precision of recommendations, a deep model for the processing of Review Texts and Rating Matrices (RTRM) is proposed. The model extracts deep-level features and combines them to make rating predictions. Then, by using the pre-trained Electra model, the implicit expression of each comment is get, and combining with the deep emotion analysis and attention mechanism, the analysis of the comment text is realized from the context semantic level. It solves the problem that it is difficult to analyze the semantics of short text; User (item) reviews interact with a rating matrix to predict the user’s rating of a product in the fusion layer module. Finally, the Mean Square Error (MSE) is used to perform performance comparison experiments on 6 sets of data sets. Experimental results show that the performance of the proposed model outperforms significantly other systems on a variety of datasets, and the average prediction error is reduced by a maximum of 12.821%, the model is suitable for recommending accurate items to users.
2022, 44(1): 254-260.
doi: 10.11999/JEIT200477
Abstract:
In order to develop fast and stable algorithm for estimating generalized eigenvector, a novel neuron-based algorithm is proposed for extracting the single generalized eigenvector. Through analyzing all of the stationary points, it is proved that the single estimation algorithm reaches the steady state if and only if the weight vector of the neural network is equal to the generalized eigenvector corresponding to the smallest generalized eigenvalue of a matrix pencil. The dynamic analysis of the single estimation algorithm is accomplished by the deterministic discrete time method and some boundary conditions are also obtained to guarantee the algorithm’s convergence. Trough applying the inflation technique, the single generalized eigenvector estimation algorithm is extended into a multiple generalized eigenvector estimation algorithm, and the number of the generalized eigenvectors can be increased according to actual requirement. Simulation experiments results prove that the proposed algorithm has good convergence, and the convergence speed is better than some existing algorithms.
In order to develop fast and stable algorithm for estimating generalized eigenvector, a novel neuron-based algorithm is proposed for extracting the single generalized eigenvector. Through analyzing all of the stationary points, it is proved that the single estimation algorithm reaches the steady state if and only if the weight vector of the neural network is equal to the generalized eigenvector corresponding to the smallest generalized eigenvalue of a matrix pencil. The dynamic analysis of the single estimation algorithm is accomplished by the deterministic discrete time method and some boundary conditions are also obtained to guarantee the algorithm’s convergence. Trough applying the inflation technique, the single generalized eigenvector estimation algorithm is extended into a multiple generalized eigenvector estimation algorithm, and the number of the generalized eigenvectors can be increased according to actual requirement. Simulation experiments results prove that the proposed algorithm has good convergence, and the convergence speed is better than some existing algorithms.
2022, 44(1): 261-270.
doi: 10.11999/JEIT200931
Abstract:
Previous techniques are not sufficient enough to deal with dehazing problems by using various hand-crafted priors and appear image hue and brightness distortion. In this paper, a saliency weighted multi-exposure fusion is proposed for single image dehazing. To produce several images with different exposures, a novel segmentation method is exploited to capture the range of global atmospheric light approximately, and a new Kirsh high-order difference filtering method is employed to optimize the transmission map. A saliency weighted multi-exposure fusion method is constructed to improve the dehazing quality. Extensive experimental results on both subjective and objective evaluation demonstrate that the proposed algorithm performs favorably against the state-of-the-art algorithms.
Previous techniques are not sufficient enough to deal with dehazing problems by using various hand-crafted priors and appear image hue and brightness distortion. In this paper, a saliency weighted multi-exposure fusion is proposed for single image dehazing. To produce several images with different exposures, a novel segmentation method is exploited to capture the range of global atmospheric light approximately, and a new Kirsh high-order difference filtering method is employed to optimize the transmission map. A saliency weighted multi-exposure fusion method is constructed to improve the dehazing quality. Extensive experimental results on both subjective and objective evaluation demonstrate that the proposed algorithm performs favorably against the state-of-the-art algorithms.
2022, 44(1): 271-278.
doi: 10.11999/JEIT200829
Abstract:
Sequences with optimal autocorrelation property have important roles in wireless communication, radar and cryptography. Therefore, in order to expand more ideal sequences that can be applied to communication systems, based on cyclotomy of order 2 and Chinese remainder theorem, three new constructions of balanced or almost balanced binary sequences of period\begin{document}$T = 4v$\end{document} ![]()
![]()
\begin{document}$v \equiv 3{\text{ }}\left( {{\rm{mod}} 4} \right)$\end{document} ![]()
![]()
\begin{document}$\left\{ {0, - 4} \right\}$\end{document} ![]()
![]()
\begin{document}$\left\{ {0, 4, - 4} \right\}$\end{document} ![]()
![]()
\begin{document}$v \equiv 1{\text{ }}\left( {{\rm{mod}} 4} \right)$\end{document} ![]()
![]()
\begin{document}$\left\{ {0, 4, - 4} \right\}$\end{document} ![]()
![]()
Sequences with optimal autocorrelation property have important roles in wireless communication, radar and cryptography. Therefore, in order to expand more ideal sequences that can be applied to communication systems, based on cyclotomy of order 2 and Chinese remainder theorem, three new constructions of balanced or almost balanced binary sequences of period
2022, 44(1): 279-287.
doi: 10.11999/JEIT200893
Abstract:
In order to defend against malicious node attacks and improve consensus efficiency in the blockchain-based medical data sharing system, a Security Consensus Algorithm of Medical Data (SCA_MD) based on credit rating is proposed. In SCA_MD, the medical block consensus model which consists of data-based nodes, consensus nodes and supervisory nodes is considered. The corresponding node identity verification mechanism is proposed to achieve rapid verification. Then, a Self-optimizing Credit Rating Division (SCRD) algorithm based on marine predators is proposed to limit the consensus power of malicious nodes. Finally, the voting mechanism and consensus mechanism of representative nodes are proposed to improve the efficiency of consensus. The experimental results show that no matter how the number of malicious nodes changes, SCA_MD can increase transaction throughput, and reduce average transaction delay and average node communication overhead.
In order to defend against malicious node attacks and improve consensus efficiency in the blockchain-based medical data sharing system, a Security Consensus Algorithm of Medical Data (SCA_MD) based on credit rating is proposed. In SCA_MD, the medical block consensus model which consists of data-based nodes, consensus nodes and supervisory nodes is considered. The corresponding node identity verification mechanism is proposed to achieve rapid verification. Then, a Self-optimizing Credit Rating Division (SCRD) algorithm based on marine predators is proposed to limit the consensus power of malicious nodes. Finally, the voting mechanism and consensus mechanism of representative nodes are proposed to improve the efficiency of consensus. The experimental results show that no matter how the number of malicious nodes changes, SCA_MD can increase transaction throughput, and reduce average transaction delay and average node communication overhead.
2022, 44(1): 288-294.
doi: 10.11999/JEIT200844
Abstract:
The emergence of the multi-controller architecture solves the scalability problem of the classic Software Defined Networking (SDN) architecture with a single centralized controller as the main control layer. In a multi-controller architecture, since the task of generating forwarding rules and filling them into the switch is delegated to the controller, the performance of the network depends largely on the placement of the controller. To reduce the total delay and balance the load among controllers, a Multi-Controller Placement Algorithm (MCPA) based on subnetting is proposed. This algorithm modifies the spectral clustering algorithm to ensure network connectivity and adds outlier processing algorithm and load balancing processing algorithm. The simulation results show that the algorithm can effectively divide the network and keep the load of each controller balanced while ensuring a low total network delay.
The emergence of the multi-controller architecture solves the scalability problem of the classic Software Defined Networking (SDN) architecture with a single centralized controller as the main control layer. In a multi-controller architecture, since the task of generating forwarding rules and filling them into the switch is delegated to the controller, the performance of the network depends largely on the placement of the controller. To reduce the total delay and balance the load among controllers, a Multi-Controller Placement Algorithm (MCPA) based on subnetting is proposed. This algorithm modifies the spectral clustering algorithm to ensure network connectivity and adds outlier processing algorithm and load balancing processing algorithm. The simulation results show that the algorithm can effectively divide the network and keep the load of each controller balanced while ensuring a low total network delay.
2022, 44(1): 295-304.
doi: 10.11999/JEIT201069
Abstract:
Through information sharing, the Internet of Vehicles (IoV) provides various applications for vehicles to improve road safety and traffic efficiency. However, the open communication between vehicles lead to vehicle privacy leakage and various attacks. Therefore, information sharing methods with security and privacy protection are very necessary, so a pairing-free and certificateless anonymous authentication scheme supporting batch authentication is proposed. In this scheme, firstly, the problem of certificate management and key escrow can be avoided by using the certificateless signature; Secondly, the combination of the long-term pseudo-identity generated by the regional authority and the short-term pseudo-identity generated by itself, the strong anonymity of vehicle and the freshness of signature are guaranteed, and the identity disclosure and the communication delay caused by Road-side-unit computing pseudo identity are avoided; Thirdly, the aggregating signature without pairing is used to provide batch verification, which reduces greatly the computational burden of RSUs in vehicle network environment; Finally, when a malicious event occurs, the Regional Trusted Authority (RTA) can track the real identity of the vehicle. Security proof and analysis show that, the scheme has high security, and meets the security requirements.
Through information sharing, the Internet of Vehicles (IoV) provides various applications for vehicles to improve road safety and traffic efficiency. However, the open communication between vehicles lead to vehicle privacy leakage and various attacks. Therefore, information sharing methods with security and privacy protection are very necessary, so a pairing-free and certificateless anonymous authentication scheme supporting batch authentication is proposed. In this scheme, firstly, the problem of certificate management and key escrow can be avoided by using the certificateless signature; Secondly, the combination of the long-term pseudo-identity generated by the regional authority and the short-term pseudo-identity generated by itself, the strong anonymity of vehicle and the freshness of signature are guaranteed, and the identity disclosure and the communication delay caused by Road-side-unit computing pseudo identity are avoided; Thirdly, the aggregating signature without pairing is used to provide batch verification, which reduces greatly the computational burden of RSUs in vehicle network environment; Finally, when a malicious event occurs, the Regional Trusted Authority (RTA) can track the real identity of the vehicle. Security proof and analysis show that, the scheme has high security, and meets the security requirements.
2022, 44(1): 305-313.
doi: 10.11999/JEIT201004
Abstract:
For the problems that the existing network traffic anomaly detection methods are not suitable for the real-time WSN (Wireless Sensor Networks) and lack reasonable decision mechanisms, a novel Wireless Sensor Networks (WSN) traffic anomaly detection scheme based on BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies) is proposed. Based on expanding the dimension of traffic characteristics, the scheme uses BIRCH algorithm to cluster traffic characteristics. By introducing the dynamic cluster threshold and neighbor cluster serial numbers, the BIRCH process is optimized to improve the clustering quality and performance robustness. Furthermore, to ensure the detection accuracy of the scheme, a comprehensive decision mechanism based on turning point is designed to detect abnormal traffic, combined with prediction and clustering results. The experimental results show that the proposed scheme has obvious advantages in detection effect and stability of detection performance.
For the problems that the existing network traffic anomaly detection methods are not suitable for the real-time WSN (Wireless Sensor Networks) and lack reasonable decision mechanisms, a novel Wireless Sensor Networks (WSN) traffic anomaly detection scheme based on BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies) is proposed. Based on expanding the dimension of traffic characteristics, the scheme uses BIRCH algorithm to cluster traffic characteristics. By introducing the dynamic cluster threshold and neighbor cluster serial numbers, the BIRCH process is optimized to improve the clustering quality and performance robustness. Furthermore, to ensure the detection accuracy of the scheme, a comprehensive decision mechanism based on turning point is designed to detect abnormal traffic, combined with prediction and clustering results. The experimental results show that the proposed scheme has obvious advantages in detection effect and stability of detection performance.
2022, 44(1): 314-323.
doi: 10.11999/JEIT201065
Abstract:
Because heterogeneous nodes of internet of vehicles have big performance difference and mobility, it leads that blockchain consensus algorithm has many problems, such as low transaction throughput and large transaction delay. Therefore, an Efficient Consistency Consensus Algorithm (ECCA) of blockchain for heterogeneous nodes in the internet of vehicles is proposed. In ECCA, the heterogeneous nodes of internet of vehicles which consists of verification nodes, general nodes and malicious nodes are considered. The credit rating mechanism is proposed to realize the classification of credit rating and division of three types of heterogeneous nodes. Then, a cross-region node identity change mechanism is proposed to adjust timely the node identity in the current region is proposed. Finally, an improved consensus algorithm is proposed to meet the timeliness requirement of the internet of vehicles. The simulation results show that the ECCA algorithm can reduce the impact of poorly performing general nodes and malicious nodes on the efficiency of block consensus, and increase transaction throughput and reduces average transaction delay and average node communication overhead.
Because heterogeneous nodes of internet of vehicles have big performance difference and mobility, it leads that blockchain consensus algorithm has many problems, such as low transaction throughput and large transaction delay. Therefore, an Efficient Consistency Consensus Algorithm (ECCA) of blockchain for heterogeneous nodes in the internet of vehicles is proposed. In ECCA, the heterogeneous nodes of internet of vehicles which consists of verification nodes, general nodes and malicious nodes are considered. The credit rating mechanism is proposed to realize the classification of credit rating and division of three types of heterogeneous nodes. Then, a cross-region node identity change mechanism is proposed to adjust timely the node identity in the current region is proposed. Finally, an improved consensus algorithm is proposed to meet the timeliness requirement of the internet of vehicles. The simulation results show that the ECCA algorithm can reduce the impact of poorly performing general nodes and malicious nodes on the efficiency of block consensus, and increase transaction throughput and reduces average transaction delay and average node communication overhead.
2022, 44(1): 324-331.
doi: 10.11999/JEIT200924
Abstract:
To overcome the policy generation problem faced by other access control mechanism in the process of migration to attribution-based access control mechanism, an access control policy generation method based on access control log is proposed. The recursive attribute elimination method is utilized to implement attribute selection. Based on information impurity, the attribute-permission relationship is extracted from the access control logs, and the result of entity attribute selection is combined to build the policy structure tree, so as to realize the policy generation of Attribute-Based Access Control (ABAC). In addition, an optimization algorithm based on binary search is designed to calculate quickly the parameters of the optimal policy generation. The experimental results show that only 32.56% of the attribute information in the original entity attribute set can be used to cover 95% of the permission in the log. The size of the policies is also reduced to 33.33% of the original size. The effectiveness of the scheme is proved.
To overcome the policy generation problem faced by other access control mechanism in the process of migration to attribution-based access control mechanism, an access control policy generation method based on access control log is proposed. The recursive attribute elimination method is utilized to implement attribute selection. Based on information impurity, the attribute-permission relationship is extracted from the access control logs, and the result of entity attribute selection is combined to build the policy structure tree, so as to realize the policy generation of Attribute-Based Access Control (ABAC). In addition, an optimization algorithm based on binary search is designed to calculate quickly the parameters of the optimal policy generation. The experimental results show that only 32.56% of the attribute information in the original entity attribute set can be used to cover 95% of the permission in the log. The size of the policies is also reduced to 33.33% of the original size. The effectiveness of the scheme is proved.
2022, 44(1): 332-338.
doi: 10.11999/JEIT200890
Abstract:
Channel modeling and simulation is the basis of performance analysis and evaluation of High frequency (HF) aviation communication system. A HF aviation mobile channel model based on Watterson model is proposed by analyzing the influence of maneuvering state parameters such as maximum moving speed, acceleration, maneuvering frequency and trajectory on the time-varying Doppler frequency shift. The model can fully describe the Doppler frequency shift and spread caused by the relative movement of the transmitter and receiver, and is suitable for HF over-the-horizon aeronautical mobile communication channel. The simulation results show that the model can realize the differential simulation of HF mobile communication channels for different types of aircraft with different maneuvering states, and can realize the customized channel simulation in specific scenarios when the route path is known.
Channel modeling and simulation is the basis of performance analysis and evaluation of High frequency (HF) aviation communication system. A HF aviation mobile channel model based on Watterson model is proposed by analyzing the influence of maneuvering state parameters such as maximum moving speed, acceleration, maneuvering frequency and trajectory on the time-varying Doppler frequency shift. The model can fully describe the Doppler frequency shift and spread caused by the relative movement of the transmitter and receiver, and is suitable for HF over-the-horizon aeronautical mobile communication channel. The simulation results show that the model can realize the differential simulation of HF mobile communication channels for different types of aircraft with different maneuvering states, and can realize the customized channel simulation in specific scenarios when the route path is known.
2022, 44(1): 339-345.
doi: 10.11999/JEIT200922
Abstract:
Satellite signal concealment technology protects the safety of signal waveform by using signals with high power and specific parameter characteristics. However, with the development of modulation recognition and Successive Interference Cancellation (SIC) technology, the cover signal using conventional modulation methods can be cracked by non-cooperative receivers. In order to meet the safety communication requirements of portable satellite communication terminals, the idea of using local signal copies for self-interference cancellation in Paired Carrier Multiple Access (PCMA) is used for reference. This paper proposes a satellite secure communication method suitable for portable terminals, and introduces Weighted-type FRactional Fourier Transform (WFRFT) as a modulation method to mask the signal, thereby increase the difficulty of non-cooperative parties in identifying and demodulating the cover signal. Simulation shows that the method proposed in this paper improves the security of signal transmission, and the loss of the system's bit error rate is within an acceptable range after signal separation, which is suitable for practical use.
Satellite signal concealment technology protects the safety of signal waveform by using signals with high power and specific parameter characteristics. However, with the development of modulation recognition and Successive Interference Cancellation (SIC) technology, the cover signal using conventional modulation methods can be cracked by non-cooperative receivers. In order to meet the safety communication requirements of portable satellite communication terminals, the idea of using local signal copies for self-interference cancellation in Paired Carrier Multiple Access (PCMA) is used for reference. This paper proposes a satellite secure communication method suitable for portable terminals, and introduces Weighted-type FRactional Fourier Transform (WFRFT) as a modulation method to mask the signal, thereby increase the difficulty of non-cooperative parties in identifying and demodulating the cover signal. Simulation shows that the method proposed in this paper improves the security of signal transmission, and the loss of the system's bit error rate is within an acceptable range after signal separation, which is suitable for practical use.
2022, 44(1): 346-353.
doi: 10.11999/JEIT200930
Abstract:
In ultra dense heterogeneous wireless network with sleep mechanism, in view of the problem that the network dynamic is enhanced and the handoff performance is reduced, a network selection algorithm based on improved deep Q-learning is proposed. Firstly, according to the dynamic analysis of the network, a deep Q-learning network selection model is constructed; Secondly, the training samples and weights of the offline training module in deep Q-learning network selection model, which are transferred to the online network decision-making module through the transfer learning; Finally, the training samples and weights of transfer are used to accelerate the process of training neural network, and the optimal network selection strategy is obtained. Experimental results demonstrate that the proposed algorithm improves significantly the performance degradation of high dynamic network handoff caused by sleep mechanism and the time complexity of traditional deep Q-learning algorithm for online network selection.
In ultra dense heterogeneous wireless network with sleep mechanism, in view of the problem that the network dynamic is enhanced and the handoff performance is reduced, a network selection algorithm based on improved deep Q-learning is proposed. Firstly, according to the dynamic analysis of the network, a deep Q-learning network selection model is constructed; Secondly, the training samples and weights of the offline training module in deep Q-learning network selection model, which are transferred to the online network decision-making module through the transfer learning; Finally, the training samples and weights of transfer are used to accelerate the process of training neural network, and the optimal network selection strategy is obtained. Experimental results demonstrate that the proposed algorithm improves significantly the performance degradation of high dynamic network handoff caused by sleep mechanism and the time complexity of traditional deep Q-learning algorithm for online network selection.
2022, 44(1): 354-362.
doi: 10.11999/JEIT201030
Abstract:
Skywave Over-The-Horizon Radar (OTHR) relies on the earth’s ionosphere which reflects its electromagnetic waves to achieve long range early warning of a variety of high-value targets. The model of ionosphere is the key factor for OTHR target tracking. Considering the spatial correlation of the Virtual Ionospheric Height (VIH) in different geographic locations, a new VIH model, represented by a Gaussian Markov Random Field (GMRF) and a new multipath target tracking method are proposed. Based on Bayesian estimation, the new method estimates jointly target state and VIHs under multipath and clutter environment. Given the limited measurements of the ionospheric subregions, the VIHs of the unmeasured subregions are inferred based on the GMRF model, thereby the precision of VIHs and the accuracy of the target tracking are improved. Numerical simulation shows that the accuracy of OTHR target tracking is improved.
Skywave Over-The-Horizon Radar (OTHR) relies on the earth’s ionosphere which reflects its electromagnetic waves to achieve long range early warning of a variety of high-value targets. The model of ionosphere is the key factor for OTHR target tracking. Considering the spatial correlation of the Virtual Ionospheric Height (VIH) in different geographic locations, a new VIH model, represented by a Gaussian Markov Random Field (GMRF) and a new multipath target tracking method are proposed. Based on Bayesian estimation, the new method estimates jointly target state and VIHs under multipath and clutter environment. Given the limited measurements of the ionospheric subregions, the VIHs of the unmeasured subregions are inferred based on the GMRF model, thereby the precision of VIHs and the accuracy of the target tracking are improved. Numerical simulation shows that the accuracy of OTHR target tracking is improved.
2022, 44(1): 363-371.
doi: 10.11999/JEIT200913
Abstract:
When there are multiple pulse train radiation source signals with unknown numbers and similar signal parameters in a given reconnaissance area, it is impossible to locate accurately multiple radiation source signals using classic multi-target localization method and Direct Position Determination (DPD) algorithm. This paper proposes a direct pulse train positioning algorithm based on the Localization information DPD (L-DPD) accumulation. The algorithm uses multiple moving reconnaissance stations to receive the signal from the fixed target radiation source, uses full the localization information of the pulse train signal and combines the time delay and frequency shift information of each pulse to locate directly the target, which solves the problem that the classic DPD algorithm can not effectively locate the time-frequency-space aliased signal. At the same time, this paper derives the Cramer-Rao Lower Bound(CRLB) of the DPD algorithm for coherent pulse signals under the background of Gaussian white noise. Simulation analysis shows that the algorithm can perform high-resolution and high-precision positioning of time-frequency-space aliased signals.
When there are multiple pulse train radiation source signals with unknown numbers and similar signal parameters in a given reconnaissance area, it is impossible to locate accurately multiple radiation source signals using classic multi-target localization method and Direct Position Determination (DPD) algorithm. This paper proposes a direct pulse train positioning algorithm based on the Localization information DPD (L-DPD) accumulation. The algorithm uses multiple moving reconnaissance stations to receive the signal from the fixed target radiation source, uses full the localization information of the pulse train signal and combines the time delay and frequency shift information of each pulse to locate directly the target, which solves the problem that the classic DPD algorithm can not effectively locate the time-frequency-space aliased signal. At the same time, this paper derives the Cramer-Rao Lower Bound(CRLB) of the DPD algorithm for coherent pulse signals under the background of Gaussian white noise. Simulation analysis shows that the algorithm can perform high-resolution and high-precision positioning of time-frequency-space aliased signals.
2022, 44(1): 372-379.
doi: 10.11999/JEIT201024
Abstract:
For the problem of a low number of consecutive lags and high redundancy of sensors in the coprime array, two sparse arrays are proposed in this paper. First, by analyzing the influence of the sensor positions on the unique lags and consecutive lags of the difference coarray, it is concluded that the range of the consecutive lags in the coprime array is not changed after removing the specific sensors. Then, entire array structure is optimized while keeping the number of sensors unchanged, increasing the number of consecutive lags. Afterward, mathematical expressions of the consecutive lags and degree of freedoms of the proposed arrays are derived respectively. Finally, simulations are carried out with the same physical sensors and the identical estimation method to verify the DOA estimation performance of the proposed sparse arrays.
For the problem of a low number of consecutive lags and high redundancy of sensors in the coprime array, two sparse arrays are proposed in this paper. First, by analyzing the influence of the sensor positions on the unique lags and consecutive lags of the difference coarray, it is concluded that the range of the consecutive lags in the coprime array is not changed after removing the specific sensors. Then, entire array structure is optimized while keeping the number of sensors unchanged, increasing the number of consecutive lags. Afterward, mathematical expressions of the consecutive lags and degree of freedoms of the proposed arrays are derived respectively. Finally, simulations are carried out with the same physical sensors and the identical estimation method to verify the DOA estimation performance of the proposed sparse arrays.
2022, 44(1): 380-389.
doi: 10.11999/JEIT201059
Abstract:
Ship targets are sparsely distributed in Synthetic Aperture Radar (SAR) images, and the design of anchor frame has a great impact on the accuracy and generalization of existing SAR image target detection method based on anchor. Therefore, an anchor-free method based on context information fusion and interacting branch for ship detection in SAR images (named as CI-Net) is proposed. Considering the diversity of ship scale in SAR images, a context fusion module is designed in the feature extraction stage, integrate high and low levels of information in a bottom-up manner and refine the extracted features to be detected by combining with the target context information. Secondly, aiming at the problem of complex targets in the scene is not accurate, interacting branch module is put forward. In the detection phase, use classification branches optimization regression testing box is used, to improve the target frame’s precision. At the same time, the new Intersection over Union (IOU) is used on branches of the classification to improve detection network classification confidence, to inhibit detection box of low quality. Experimental results show that the proposed method achieves good detection results on both SSDD and SAR-Ship-Dataset, with Average Precision (AP) reaching 92.56% and 88.32%, respectively. Compared with other ship detection methods in SAR image, the proposed method not only has excellent performance in accuracy, but also has a faster detection speed after abandoning the complex calculation related to anchor frame. It also has a certain practical significance for real-time target detection in SAR image.
Ship targets are sparsely distributed in Synthetic Aperture Radar (SAR) images, and the design of anchor frame has a great impact on the accuracy and generalization of existing SAR image target detection method based on anchor. Therefore, an anchor-free method based on context information fusion and interacting branch for ship detection in SAR images (named as CI-Net) is proposed. Considering the diversity of ship scale in SAR images, a context fusion module is designed in the feature extraction stage, integrate high and low levels of information in a bottom-up manner and refine the extracted features to be detected by combining with the target context information. Secondly, aiming at the problem of complex targets in the scene is not accurate, interacting branch module is put forward. In the detection phase, use classification branches optimization regression testing box is used, to improve the target frame’s precision. At the same time, the new Intersection over Union (IOU) is used on branches of the classification to improve detection network classification confidence, to inhibit detection box of low quality. Experimental results show that the proposed method achieves good detection results on both SSDD and SAR-Ship-Dataset, with Average Precision (AP) reaching 92.56% and 88.32%, respectively. Compared with other ship detection methods in SAR image, the proposed method not only has excellent performance in accuracy, but also has a faster detection speed after abandoning the complex calculation related to anchor frame. It also has a certain practical significance for real-time target detection in SAR image.
2022, 44(1): 390-399.
doi: 10.11999/JEIT201095
Abstract:
A new four-dimensional chaotic system with extreme multi-stability based on a classic three-dimensional chaotic system is proposed. The new system has a line equilibrium point, which can generate an infinite number of symmetrical homogeneous attractors. The chaotic characteristics of the system are analyzed by phase orbit diagram and Poincaré section methods. Using phase orbit diagrams, bifurcation diagrams and Lyapunov exponent spectrum methods, the influence of initial conditions on the extreme multi-stability of the system is analyzed. The analysis shows that the system has a large initial value variation range, and the Lyapunov exponent spectrum is constant except for the zero point. In addition, the system also has centrally symmetrical discrete bifurcation diagrams. Furthermore, the relationship between the initial symmetry of the system and the symmetry of the attractor is studied, which is different from the symmetrical attractor in the existing chaotic system, which can generate an infinite number of symmetrical homogeneous attractors. Finally, circuit simulation software is used to build an analog circuit to capture the chaotic attractor of the system, and the result verifies the correctness of the numerical simulation.
A new four-dimensional chaotic system with extreme multi-stability based on a classic three-dimensional chaotic system is proposed. The new system has a line equilibrium point, which can generate an infinite number of symmetrical homogeneous attractors. The chaotic characteristics of the system are analyzed by phase orbit diagram and Poincaré section methods. Using phase orbit diagrams, bifurcation diagrams and Lyapunov exponent spectrum methods, the influence of initial conditions on the extreme multi-stability of the system is analyzed. The analysis shows that the system has a large initial value variation range, and the Lyapunov exponent spectrum is constant except for the zero point. In addition, the system also has centrally symmetrical discrete bifurcation diagrams. Furthermore, the relationship between the initial symmetry of the system and the symmetry of the attractor is studied, which is different from the symmetrical attractor in the existing chaotic system, which can generate an infinite number of symmetrical homogeneous attractors. Finally, circuit simulation software is used to build an analog circuit to capture the chaotic attractor of the system, and the result verifies the correctness of the numerical simulation.
Design and Implementation of Memristor-based Chaotic Synchronization under a Single Input Controller
2022, 44(1): 400-407.
doi: 10.11999/JEIT200947
Abstract:
A memristor-based chaotic synchronization circuit is designed and implemented under a single-input controller, and it is applied to secure communication based on memristor chaotic synchronization. Firstly, based on the chaotic synchronization theory, a chaotic synchronization system and secure communication model are constructed, and a fourth-order memristor-based chaotic circuit is implemented, the chaotic en-/decryption circuit is also designed. Secondly, the proposed memristor-based chaotic circuit is considered as the chaotic drive and response circuits, and a single-input chaotic synchronization controller is designed according to their error system, and it is implemented in the memristor-based chaotic synchronization circuit. Finally, many experiments of chaotic synchronization and secure communication based on chaotic synchronization are performed, and experimental results show that the proposed memristor-based chaotic synchronization circuit has many advantages, such as simple structure, convenient operation and good waveform. Furthermore, secure communication based on chaotic synchronization has high signal recovery capability and good anti-decipher ability, so that it has certain theoretical significance and potential practical value.
A memristor-based chaotic synchronization circuit is designed and implemented under a single-input controller, and it is applied to secure communication based on memristor chaotic synchronization. Firstly, based on the chaotic synchronization theory, a chaotic synchronization system and secure communication model are constructed, and a fourth-order memristor-based chaotic circuit is implemented, the chaotic en-/decryption circuit is also designed. Secondly, the proposed memristor-based chaotic circuit is considered as the chaotic drive and response circuits, and a single-input chaotic synchronization controller is designed according to their error system, and it is implemented in the memristor-based chaotic synchronization circuit. Finally, many experiments of chaotic synchronization and secure communication based on chaotic synchronization are performed, and experimental results show that the proposed memristor-based chaotic synchronization circuit has many advantages, such as simple structure, convenient operation and good waveform. Furthermore, secure communication based on chaotic synchronization has high signal recovery capability and good anti-decipher ability, so that it has certain theoretical significance and potential practical value.
2022, 44(1): 1-10.
doi: 10.11999/JEIT210879
Abstract:
As a typical medical robot, the efficiency of ultrasound imaging and the fatigue caused by manual operation for a long time in assisted diagnosis and surgical guidance can effectively be reduced by ultrasound robots. To improve the imaging efficiency and stability of ultrasound robots in complex dynamic environments, a deep reinforcement learning-based imaging control method and system are proposed. Firstly, an imaging action decision method based on proximal policy gradient optimization is proposed to generate spatial action and probe pose motion decisions of the ultrasound robot in real-time and to realize the continuous generation process of imaging action decisions for targets in dynamic environments. Further, based on the characteristics of the complex and flexible environment faced by the ultrasound robot in the imaging task, an ultrasound robot control optimization strategy is proposed on the basis of the autonomous ultrasound robot motion decision. Eventually, a fully autonomous robotic ultrasound imaging process for different human body parts is achieved while avoiding parameter adjustments and complex dynamic environments.
As a typical medical robot, the efficiency of ultrasound imaging and the fatigue caused by manual operation for a long time in assisted diagnosis and surgical guidance can effectively be reduced by ultrasound robots. To improve the imaging efficiency and stability of ultrasound robots in complex dynamic environments, a deep reinforcement learning-based imaging control method and system are proposed. Firstly, an imaging action decision method based on proximal policy gradient optimization is proposed to generate spatial action and probe pose motion decisions of the ultrasound robot in real-time and to realize the continuous generation process of imaging action decisions for targets in dynamic environments. Further, based on the characteristics of the complex and flexible environment faced by the ultrasound robot in the imaging task, an ultrasound robot control optimization strategy is proposed on the basis of the autonomous ultrasound robot motion decision. Eventually, a fully autonomous robotic ultrasound imaging process for different human body parts is achieved while avoiding parameter adjustments and complex dynamic environments.
2022, 44(1): 11-17.
doi: 10.11999/JEIT210710
Abstract:
Most of the existing multi-modal segmentation methods are adopted on the co-registered multi-modal images. However, these two-stage algorithms of the segmentation and the registration achieve low segmentation performance on the modalities with remarkable spatial misalignment. To solve this problem, a cross-modal Spatial Alignment based Multi-Modal pulmonary mass Segmentation Network (MMSASegNet) with low model complexity and high segmentation accuracy is proposed. Dual-path Res-UNet is adopted as the backbone segmentation architecture of the proposed network for the better multi-modal feature extraction. Spatial Transformer Networks (STN) is applied to the segmentation masks from two paths to align the spatial information of mass region. In order to realize the multi-modal feature fusion based on the spatial alignment on the region of mass, the deformed mask and the reference mask are matrix-multiplied by the feature maps of each modality respectively. Further, the yielding cross-modality spatially aligned feature maps from multiple modalities are fused and learned through the feature fusion module for the multi-modal mass segmentation. In order to improve the performance of the end-to-end multi-modal segmentation network, deep supervision learning strategy is employed with the joint cost function constraining mass segmentation, mass spatial alignment and feature fusion. Moreover, the multi-stage training strategy is adopted to improve the training efficiency of each module. On the pulmonary mass datasets containing T2-Weighted-MRI（T2W） and Diffusion-Weighted-MRI Images（DWI）, the proposed method achieved improvement on the metrics of Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD).
Most of the existing multi-modal segmentation methods are adopted on the co-registered multi-modal images. However, these two-stage algorithms of the segmentation and the registration achieve low segmentation performance on the modalities with remarkable spatial misalignment. To solve this problem, a cross-modal Spatial Alignment based Multi-Modal pulmonary mass Segmentation Network (MMSASegNet) with low model complexity and high segmentation accuracy is proposed. Dual-path Res-UNet is adopted as the backbone segmentation architecture of the proposed network for the better multi-modal feature extraction. Spatial Transformer Networks (STN) is applied to the segmentation masks from two paths to align the spatial information of mass region. In order to realize the multi-modal feature fusion based on the spatial alignment on the region of mass, the deformed mask and the reference mask are matrix-multiplied by the feature maps of each modality respectively. Further, the yielding cross-modality spatially aligned feature maps from multiple modalities are fused and learned through the feature fusion module for the multi-modal mass segmentation. In order to improve the performance of the end-to-end multi-modal segmentation network, deep supervision learning strategy is employed with the joint cost function constraining mass segmentation, mass spatial alignment and feature fusion. Moreover, the multi-stage training strategy is adopted to improve the training efficiency of each module. On the pulmonary mass datasets containing T2-Weighted-MRI（T2W） and Diffusion-Weighted-MRI Images（DWI）, the proposed method achieved improvement on the metrics of Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD).
2022, 44(1): 18-26.
doi: 10.11999/JEIT210919
Abstract:
In vitrectomy combined with silicone oil tamponade for the treatment of Rhegmatogenous Retinal Detachment(RRD) in ophthalmology, the prediction of postoperative silicone oil emulsification, the appropriate amount of silicone oil filling and the final removal time are the key to the success of the operation. However, doctors can not directly observe the internal structure of the eyeball during the operation process, and it is difficult to analyze quantitatively the filling materials. After the operation, there is also a lack of visual cognition of the emulsification state of silicone oil. In this paper, a volume incompressible multiphase fluid computational framework based on Smooth Particle Hydrodynamics is proposed to calculate numerically the coupling of silicone oil and water in intraocular environment, combined with the surface tension model. A multiphase fluid miscible diffusion simulation method based on the force-balanced dispersion dynamics model is constructed to analyze visually and quantitatively the silicone oil emulsification process. Experiments show that this method can simulate the multiphase fluid coupling under strong surface tension and interphase interaction of miscible multiphase fluid stably, so as to assist effectively the doctors to determine the amount of silicone oil required before operation, and predict and evaluate the influence of silicone oil emulsification state on the prognosis of operation.
In vitrectomy combined with silicone oil tamponade for the treatment of Rhegmatogenous Retinal Detachment(RRD) in ophthalmology, the prediction of postoperative silicone oil emulsification, the appropriate amount of silicone oil filling and the final removal time are the key to the success of the operation. However, doctors can not directly observe the internal structure of the eyeball during the operation process, and it is difficult to analyze quantitatively the filling materials. After the operation, there is also a lack of visual cognition of the emulsification state of silicone oil. In this paper, a volume incompressible multiphase fluid computational framework based on Smooth Particle Hydrodynamics is proposed to calculate numerically the coupling of silicone oil and water in intraocular environment, combined with the surface tension model. A multiphase fluid miscible diffusion simulation method based on the force-balanced dispersion dynamics model is constructed to analyze visually and quantitatively the silicone oil emulsification process. Experiments show that this method can simulate the multiphase fluid coupling under strong surface tension and interphase interaction of miscible multiphase fluid stably, so as to assist effectively the doctors to determine the amount of silicone oil required before operation, and predict and evaluate the influence of silicone oil emulsification state on the prognosis of operation.
2022, 44(1): 27-38.
doi: 10.11999/JEIT210931
Abstract:
The development of Acute Kidney Injury (AKI) during admission to the Intensive Care Unit (ICU) is associated with increased morbidity and mortality. The objective of this study is to develop a machine learning-based framework for interpretable AKI prediction in critical care that can achieve both good prediction and interpretation capability. Data extracted from the Medical Information Mart for Intensive Care III (MIMIC-III) include patient age, gender, vital signs and lab values during the first day of ICU admission and subsequent hospitalization. In this study, the prediction performance of the XGBoost model is demonstrated by comparing it to four other machine learning models. In addition, the SHapley Additive exPlanation (SHAP) framework is used to provide individualized evaluation and explanations to enable personalized clinical decision support. The results show that XGBoost can predict AKI robustly with an Accuracy and the area Under the receiver operating Characteristic curve (AUC) of 0.824 and 0.840, respectively, which are higher than previous prediction models. Furthermore, a feature dependency analysis is conducted for two pairs of features and found decrease in urine volume and elevation of blood urea nitrogen indicates an increase of AKI risk. To sum up, this interpretable predictive model may help clinicians more accurately identify patients at risk of AKI in intensive care and provide better treatment for patients. In addition, the use of this interpretability framework increases model transparency and facilitates clinicians to analyze the reliability of predictive models.
The development of Acute Kidney Injury (AKI) during admission to the Intensive Care Unit (ICU) is associated with increased morbidity and mortality. The objective of this study is to develop a machine learning-based framework for interpretable AKI prediction in critical care that can achieve both good prediction and interpretation capability. Data extracted from the Medical Information Mart for Intensive Care III (MIMIC-III) include patient age, gender, vital signs and lab values during the first day of ICU admission and subsequent hospitalization. In this study, the prediction performance of the XGBoost model is demonstrated by comparing it to four other machine learning models. In addition, the SHapley Additive exPlanation (SHAP) framework is used to provide individualized evaluation and explanations to enable personalized clinical decision support. The results show that XGBoost can predict AKI robustly with an Accuracy and the area Under the receiver operating Characteristic curve (AUC) of 0.824 and 0.840, respectively, which are higher than previous prediction models. Furthermore, a feature dependency analysis is conducted for two pairs of features and found decrease in urine volume and elevation of blood urea nitrogen indicates an increase of AKI risk. To sum up, this interpretable predictive model may help clinicians more accurately identify patients at risk of AKI in intensive care and provide better treatment for patients. In addition, the use of this interpretability framework increases model transparency and facilitates clinicians to analyze the reliability of predictive models.
2022, 44(1): 39-47.
doi: 10.11999/JEIT210916
Abstract:
The precise segmentation of colon polyps plays a significant role in the diagnosis and treatment of colorectal cancer. The existing segmentation methods have generally artifacts and low segmentation accuracy. In this paper, Stair-structured U-Net (SU-Net) is proposed to segment polyp, using U-shaped structure. The Kronecker product is used to extend the standard atrous convolution kernel to keep more detail structrural features that are easily ignored. Stair-structured fusion module is applied to encompass effectively multi-scale features. The decoder introduces a convolutional reshaped upsampling module to generate pixel-level predictions. Experiments are performed on the Kvasir-SEG dataset and the CVC-EndoSceneStill dataset. The results show that the method proposed in this paper outperforms other polyp segmentation methods in Dice and Intersection-over-Union(IoU).
The precise segmentation of colon polyps plays a significant role in the diagnosis and treatment of colorectal cancer. The existing segmentation methods have generally artifacts and low segmentation accuracy. In this paper, Stair-structured U-Net (SU-Net) is proposed to segment polyp, using U-shaped structure. The Kronecker product is used to extend the standard atrous convolution kernel to keep more detail structrural features that are easily ignored. Stair-structured fusion module is applied to encompass effectively multi-scale features. The decoder introduces a convolutional reshaped upsampling module to generate pixel-level predictions. Experiments are performed on the Kvasir-SEG dataset and the CVC-EndoSceneStill dataset. The results show that the method proposed in this paper outperforms other polyp segmentation methods in Dice and Intersection-over-Union(IoU).
2022, 44(1): 48-58.
doi: 10.11999/JEIT210917
Abstract:
Since the outbreak of the Covid-19 epidemic in the world in late 2019, all countries in the world are under the threat of epidemic. Covid-19 invades the body's respiratory system, causing lung infection or even death. Computed Tomography (CT) is a routine method for doctors to diagnose patients with pneumonia. In order to improve the efficiency of doctors in diagnosing patients with new crown infection, this paper proposes a semantic segmentation network LRSAR-Net based on low rank tensor self-attention reconstruction, in which the low rank tensor self-attention reconstruction module is used to obtain long-range information. The low rank tensor self-attention reconstruction module mainly includes three parts: low rank tensor generation sub module, low rank self-attention sub module and high rank tensor reconstruction module. The low rank tensor self-attention module is divided into multiple low rank tensors, the low rank self-attention feature map is constructed, and then the multiple low rank tensor attention feature maps are reconstructed into a high rank attention feature map. The self-attention module obtains long-range semantic information by calculating the similarity matrix. Compared with the traditional self-attention module Non Local, the low rank tensor self-attention reconstruction module has lower computational complexity and faster computing speed. Finally, this paper compares with other excellent semantic segmentation models to reflect the effectiveness of the model.
Since the outbreak of the Covid-19 epidemic in the world in late 2019, all countries in the world are under the threat of epidemic. Covid-19 invades the body's respiratory system, causing lung infection or even death. Computed Tomography (CT) is a routine method for doctors to diagnose patients with pneumonia. In order to improve the efficiency of doctors in diagnosing patients with new crown infection, this paper proposes a semantic segmentation network LRSAR-Net based on low rank tensor self-attention reconstruction, in which the low rank tensor self-attention reconstruction module is used to obtain long-range information. The low rank tensor self-attention reconstruction module mainly includes three parts: low rank tensor generation sub module, low rank self-attention sub module and high rank tensor reconstruction module. The low rank tensor self-attention module is divided into multiple low rank tensors, the low rank self-attention feature map is constructed, and then the multiple low rank tensor attention feature maps are reconstructed into a high rank attention feature map. The self-attention module obtains long-range semantic information by calculating the similarity matrix. Compared with the traditional self-attention module Non Local, the low rank tensor self-attention reconstruction module has lower computational complexity and faster computing speed. Finally, this paper compares with other excellent semantic segmentation models to reflect the effectiveness of the model.
2022, 44(1): 59-69.
doi: 10.11999/JEIT210922
Abstract:
BallistoCardioGraphy (BCG) signal contains physiological parameters during sleep for example heartbeat. It is measured by non-contact method, therefore its application is limited due to interference. ElectroCardioGram (ECG) signals are widely used, but it is difficult to operate using contact measure. In order to realize the non-contact measurement and monitoring of ECG signals, this paper introduces the Parameterless Scale space Approach (PSA) and improves the Empirical Wavelet Transform (EWT) algorithm to decompose the heartbeat component from BCG signal. The results show that the proposed method can effectively decompose the heartbeat signal from BCG signal to the maximum extent. On this basis, ECG signals are reconstructed by improved Deep Convolutional Generative Adversarial Networks (DCGAN). The results show that the ECG signals can be reconstructed from heartbeat components by the proposed algorithm, and the root mean square error is –16.8422 dB.
BallistoCardioGraphy (BCG) signal contains physiological parameters during sleep for example heartbeat. It is measured by non-contact method, therefore its application is limited due to interference. ElectroCardioGram (ECG) signals are widely used, but it is difficult to operate using contact measure. In order to realize the non-contact measurement and monitoring of ECG signals, this paper introduces the Parameterless Scale space Approach (PSA) and improves the Empirical Wavelet Transform (EWT) algorithm to decompose the heartbeat component from BCG signal. The results show that the proposed method can effectively decompose the heartbeat signal from BCG signal to the maximum extent. On this basis, ECG signals are reconstructed by improved Deep Convolutional Generative Adversarial Networks (DCGAN). The results show that the ECG signals can be reconstructed from heartbeat components by the proposed algorithm, and the root mean square error is –16.8422 dB.
2022, 44(1): 70-77.
doi: 10.11999/JEIT210920
Abstract:
Gastrointestinal endoscopy plays a critical role in examination and diagnosis upper gastrointestinal diseases. The motion blur of endoscopic images can interfere with doctor's judgment and machine-assisted diagnosis. Due to the lack of attention to structural information in existing deblurring networks, artifacts and structural distortions occur easily when processing endoscopic images. In order to solve this problem and improve the image quality of gastroscopy, a gradient-guided generative adversarial network is proposed in this paper. The network uses the Res2net structure as the backbone, including two interactive branches, the image branch with its intensity and the gradient one. The gradient branch guides the deblurring and reconstruction of the image which in the other branch. Thus more structure information of the image can be kept, with less artifacts and alleviating structural deformation. A quasi-lightweight preprocessing network is designed to correct excessive blur and improve training efficiency. Experiments are performed on the traditional gastroscopy and the capsule gastroscopy datasets. The test results show that the Peak Signal to Noise Ratio(PSNR) and Structural SIMilarity(SSIM) indicators of the algorithm are better than those of the comparison algorithms, and the visual effect of the restored image is evidently improved, without obvious artifacts and structural deformation.
Gastrointestinal endoscopy plays a critical role in examination and diagnosis upper gastrointestinal diseases. The motion blur of endoscopic images can interfere with doctor's judgment and machine-assisted diagnosis. Due to the lack of attention to structural information in existing deblurring networks, artifacts and structural distortions occur easily when processing endoscopic images. In order to solve this problem and improve the image quality of gastroscopy, a gradient-guided generative adversarial network is proposed in this paper. The network uses the Res2net structure as the backbone, including two interactive branches, the image branch with its intensity and the gradient one. The gradient branch guides the deblurring and reconstruction of the image which in the other branch. Thus more structure information of the image can be kept, with less artifacts and alleviating structural deformation. A quasi-lightweight preprocessing network is designed to correct excessive blur and improve training efficiency. Experiments are performed on the traditional gastroscopy and the capsule gastroscopy datasets. The test results show that the Peak Signal to Noise Ratio(PSNR) and Structural SIMilarity(SSIM) indicators of the algorithm are better than those of the comparison algorithms, and the visual effect of the restored image is evidently improved, without obvious artifacts and structural deformation.
2022, 44(1): 78-88.
doi: 10.11999/JEIT210909
Abstract:
As the gold standard for the detection of Esophageal Motility Disorder(EMD), High-Resolution Manometry(HRM) is widely used in clinical tests to assist doctors in diagnosis. The amount of HRM images explodes with an increase in the prevalence rate, and the diagnostic process of EMD is complicated, both of which may lead to misdiagnosis of EMD in clinic. To improve the accuracy of the diagnosis of EMD, we hope to build a Computer Aided Diagnosis(CAD) system to assist doctors in analyzing HRM images automatically. Since the abnormality of esophageal contraction vigor is an important basis for diagnosis of EMD, in this paper, a Deep Learning(DL) model(PoS-ClasNet) is proposed to classify esophageal contraction vigor, which lays the foundation for machine to diagnose EMD instead of manual in the future. PoS-ClasNet, as a multi-task Convolutional Neural Network(CNN), is formed by PoSNet and S-ClassNet. The former is used to detect and extract swallowing frames in HRM images, while the latter identifies the type of contraction vigor based on esophageal swallowing characteristics. 4,000 expert-labeled HRM images are used for the experiment, among which the images of training set, verification set and test set accounted for 70%, 20% and 10%. On the test set, the classification accuracy of esophageal contraction vigor classifier PoS-ClasNet is as high as 93.25%, meanwhile the precision rate and the recall rate are 93.39% and 93.60% respectively. The experimental results show PoS-ClasNet can well adapt to the features of HRM image, with the outstanding accuracy and robustness in the task of intelligent diagnosis of esophageal contraction vigor. If the proposed model is used to assist doctors in clinical prevention, diagnosis and treatment, it will bring enormous social benefits.
As the gold standard for the detection of Esophageal Motility Disorder(EMD), High-Resolution Manometry(HRM) is widely used in clinical tests to assist doctors in diagnosis. The amount of HRM images explodes with an increase in the prevalence rate, and the diagnostic process of EMD is complicated, both of which may lead to misdiagnosis of EMD in clinic. To improve the accuracy of the diagnosis of EMD, we hope to build a Computer Aided Diagnosis(CAD) system to assist doctors in analyzing HRM images automatically. Since the abnormality of esophageal contraction vigor is an important basis for diagnosis of EMD, in this paper, a Deep Learning(DL) model(PoS-ClasNet) is proposed to classify esophageal contraction vigor, which lays the foundation for machine to diagnose EMD instead of manual in the future. PoS-ClasNet, as a multi-task Convolutional Neural Network(CNN), is formed by PoSNet and S-ClassNet. The former is used to detect and extract swallowing frames in HRM images, while the latter identifies the type of contraction vigor based on esophageal swallowing characteristics. 4,000 expert-labeled HRM images are used for the experiment, among which the images of training set, verification set and test set accounted for 70%, 20% and 10%. On the test set, the classification accuracy of esophageal contraction vigor classifier PoS-ClasNet is as high as 93.25%, meanwhile the precision rate and the recall rate are 93.39% and 93.60% respectively. The experimental results show PoS-ClasNet can well adapt to the features of HRM image, with the outstanding accuracy and robustness in the task of intelligent diagnosis of esophageal contraction vigor. If the proposed model is used to assist doctors in clinical prevention, diagnosis and treatment, it will bring enormous social benefits.
2022, 44(1): 89-98.
doi: 10.11999/JEIT210935
Abstract:
Tongue color is one of the most concerned diagnostic features of tongue diagnosis in Traditional Chinese Medicine (TCM). Automatic and accurate tongue color classification is an important content of the objectification of tongue diagnosis. Due to the vagueness of the visual boundaries between different types of tongue colors and the subjectivity of the doctors, the annotated tongue image data samples often contain noises, which has a negative effect on the training of the tongue color classification model. Therefore, in this paper, a tongue color classification method in TCM with noisy labels is proposed. Firstly, a two-stage data cleaning method is proposed to identify and clean noisy labeled samples. Secondly, a lightweight Convolutional Neural Network (CNN) based on the channel attention mechanism is designed in this paper to achieve accurate classification of tongue color by enhancing the expressiveness of features. Finally, a knowledge distillation strategy with a noise sample filtering mechanism is proposed. This strategy adds a noise sample filtering mechanism led by the teacher network to eliminate further noise samples. At the same time, the teacher network is used to guide the training of the light convolutional neural network to improve the classification performance.The experimental results on the self-established TCM tongue color classification dataset show that the proposed method in this paper can significantly improve the classification accuracy with lower computational complexity, reaching 93.88%.
Tongue color is one of the most concerned diagnostic features of tongue diagnosis in Traditional Chinese Medicine (TCM). Automatic and accurate tongue color classification is an important content of the objectification of tongue diagnosis. Due to the vagueness of the visual boundaries between different types of tongue colors and the subjectivity of the doctors, the annotated tongue image data samples often contain noises, which has a negative effect on the training of the tongue color classification model. Therefore, in this paper, a tongue color classification method in TCM with noisy labels is proposed. Firstly, a two-stage data cleaning method is proposed to identify and clean noisy labeled samples. Secondly, a lightweight Convolutional Neural Network (CNN) based on the channel attention mechanism is designed in this paper to achieve accurate classification of tongue color by enhancing the expressiveness of features. Finally, a knowledge distillation strategy with a noise sample filtering mechanism is proposed. This strategy adds a noise sample filtering mechanism led by the teacher network to eliminate further noise samples. At the same time, the teacher network is used to guide the training of the light convolutional neural network to improve the classification performance.The experimental results on the self-established TCM tongue color classification dataset show that the proposed method in this paper can significantly improve the classification accuracy with lower computational complexity, reaching 93.88%.
Atrial Fibrillation Detection Based on Hilbert-Huang Transform and Deep Convolutional Neural Network
2022, 44(1): 99-106.
doi: 10.11999/JEIT211171
Abstract:
Atrial fibrillation is a common arrhythmia and its morbidity increases with age. Thus, stroke risk and cardiogenic mortality can be significantly reduced by early atrial fibrillation detection from ElectroCardioGram (ECG). In order to improve effectively detection accuracy, a novel approach is proposed to detect atrial fibrillation based on Hilbert-Huang Transform(HHT) and deep convolutional neural network. HHT is employed to transform electrocardiogram from time domain to time-frequency domain so as to enrich the feature of original data. Following that, DenseNet is introduced to deal with the detailed graph and the best model is selected during the iteration. The optimal model obtained by the proposed method achieves 99.11% and 97.25% accuracy respectively on the Massachusetts Institute of Technology - Beth Israel Hospital(MIT-BIH) and 2017 PhysioNet Challenge atrial fibrillation databases. In addition, HHT and DenseNet are compared with other time-frequency analysis and convolutional neural networks, respectively. Compared with some existing methods, the results proved that atrial fibrillation detection by HHT and Deep Convolutional Neural Network(DCNN) obtains a high detection performance.
Atrial fibrillation is a common arrhythmia and its morbidity increases with age. Thus, stroke risk and cardiogenic mortality can be significantly reduced by early atrial fibrillation detection from ElectroCardioGram (ECG). In order to improve effectively detection accuracy, a novel approach is proposed to detect atrial fibrillation based on Hilbert-Huang Transform(HHT) and deep convolutional neural network. HHT is employed to transform electrocardiogram from time domain to time-frequency domain so as to enrich the feature of original data. Following that, DenseNet is introduced to deal with the detailed graph and the best model is selected during the iteration. The optimal model obtained by the proposed method achieves 99.11% and 97.25% accuracy respectively on the Massachusetts Institute of Technology - Beth Israel Hospital(MIT-BIH) and 2017 PhysioNet Challenge atrial fibrillation databases. In addition, HHT and DenseNet are compared with other time-frequency analysis and convolutional neural networks, respectively. Compared with some existing methods, the results proved that atrial fibrillation detection by HHT and Deep Convolutional Neural Network(DCNN) obtains a high detection performance.
2022, 44(1): 107-117.
doi: 10.11999/JEIT210858
Abstract:
In order to achieve multi-user data search in electronic medical record system, an attribute based searchable encryption scheme is proposed. In this scheme, ciphertext and secure indexes are stored in the medical cloud. When the users want to access the medical data, the attribute-base searchable encryption algorithm is used for data search, and the fine-grained access control is realized. At the same time, ciphertext verification algorithm is introduced into the scheme, it solves the problem of incorrect search results under the semi-honest and curious cloud server models. The scheme uses data deduplication technology to eliminate duplicate data and reduce the storage space of the medical cloud. The scheme also realizes the hiding of access policy, and the privacy security of data users is guaranteed. The security analysis shows that the proposed scheme can well protect the confidentiality of medical data and the anonymity of users. The performance analysis demonstrate that the proposed scheme has better performance; Hence, it is more suitable for many-to-many application scenarios such as smart healthcare. It effectively realizes the sharing of patient electronic medical records between doctors and third-party users without infringing on patient privacy.
In order to achieve multi-user data search in electronic medical record system, an attribute based searchable encryption scheme is proposed. In this scheme, ciphertext and secure indexes are stored in the medical cloud. When the users want to access the medical data, the attribute-base searchable encryption algorithm is used for data search, and the fine-grained access control is realized. At the same time, ciphertext verification algorithm is introduced into the scheme, it solves the problem of incorrect search results under the semi-honest and curious cloud server models. The scheme uses data deduplication technology to eliminate duplicate data and reduce the storage space of the medical cloud. The scheme also realizes the hiding of access policy, and the privacy security of data users is guaranteed. The security analysis shows that the proposed scheme can well protect the confidentiality of medical data and the anonymity of users. The performance analysis demonstrate that the proposed scheme has better performance; Hence, it is more suitable for many-to-many application scenarios such as smart healthcare. It effectively realizes the sharing of patient electronic medical records between doctors and third-party users without infringing on patient privacy.
2022, 44(1): 118-126.
doi: 10.11999/JEIT210926
Abstract:
Medical machine translation is of great value for cross-border medical translation. Chinese to English neural machine translation has made great progress based on deep learning, powerful modeling ability and large-scale bilingual parallel data. Neural machine translation relies usually on large-scale parallel sentence pairs to train translation models. At present, Chinese-English translation data are mainly in the fields of news, policy and so on. Due to the lack of parallel data in the medical field, the performance of Chinese to English machine translation in the medical field is not compromising. To reduce the size of parallel data for training medical machine translation models, this paper proposes a paraphrase based data augmentation mechanism. The experimental results on a variety of neural machine translation models show that data augmentation through paraphrase augmentation can effectively improve the performance of medical machine translation, and has achieved consistency improvements on mainstream models such as RNNSearch and Transformers, which verifies the effectiveness of paraphrase augmentation method for domain machine translation. Meanwhile, the medical machine translation performances could be further improved based on large-scale pre-training language model, such as MT5.
Medical machine translation is of great value for cross-border medical translation. Chinese to English neural machine translation has made great progress based on deep learning, powerful modeling ability and large-scale bilingual parallel data. Neural machine translation relies usually on large-scale parallel sentence pairs to train translation models. At present, Chinese-English translation data are mainly in the fields of news, policy and so on. Due to the lack of parallel data in the medical field, the performance of Chinese to English machine translation in the medical field is not compromising. To reduce the size of parallel data for training medical machine translation models, this paper proposes a paraphrase based data augmentation mechanism. The experimental results on a variety of neural machine translation models show that data augmentation through paraphrase augmentation can effectively improve the performance of medical machine translation, and has achieved consistency improvements on mainstream models such as RNNSearch and Transformers, which verifies the effectiveness of paraphrase augmentation method for domain machine translation. Meanwhile, the medical machine translation performances could be further improved based on large-scale pre-training language model, such as MT5.
2022, 44(1): 127-137.
doi: 10.11999/JEIT200996
Abstract:
In view of the problem of low segmentation accuracy caused by the multi-scale of the lesion location in Computed-Tomography (CT) images of cerebral hemorrhage, an image segmentation model based on Attention improved U-shaped neural Network plus (AU-Net+) is proposed. Firstly, the model uses the encoder in U-Net to encode the features of the CT image of cerebral hemorrhage, and applies the proposed Residual Octave Convolution (ROC) block to the jump connection part of the U-shaped neural network to make the features of different levels more blend well. Secondly, for the merged features, a hybrid attention mechanism is introduced to improve the feature extraction ability of the target area. Finally, the Dice loss function is improved to enhance further the feature learning of the model for small and medium-sized target regions in CT images of cerebral hemorrhage. To verify the performance of the model, the mIoU index is improved by 20.9%, 3.6%, 7.0%, 3.1% compared with U-Net, Attention U-Net, UNet++ and CE-Net respectively, which indicates that AU-Net+ model has better segmentation effect.
In view of the problem of low segmentation accuracy caused by the multi-scale of the lesion location in Computed-Tomography (CT) images of cerebral hemorrhage, an image segmentation model based on Attention improved U-shaped neural Network plus (AU-Net+) is proposed. Firstly, the model uses the encoder in U-Net to encode the features of the CT image of cerebral hemorrhage, and applies the proposed Residual Octave Convolution (ROC) block to the jump connection part of the U-shaped neural network to make the features of different levels more blend well. Secondly, for the merged features, a hybrid attention mechanism is introduced to improve the feature extraction ability of the target area. Finally, the Dice loss function is improved to enhance further the feature learning of the model for small and medium-sized target regions in CT images of cerebral hemorrhage. To verify the performance of the model, the mIoU index is improved by 20.9%, 3.6%, 7.0%, 3.1% compared with U-Net, Attention U-Net, UNet++ and CE-Net respectively, which indicates that AU-Net+ model has better segmentation effect.
2022, 44(1): 138-148.
doi: 10.11999/JEIT210900
Abstract:
In order to determine accurately International Society for Urology and Pathology (ISUP) grade of clear cell Renal Cell Carcinoma (ccRCC) and achieve subsequently better treatment and prognosis, a novel channel attention mechanism named sECANet is proposed. To obtain more useful features from the feature map, sECANet calculates the information interaction of the current channel and local channels, and calculates additionally the interaction of the current channel and remote channels. A total of 90 pathological images are collected and subsequently cut into patches. After data augmentation, 5-fold cross-validation is employed to verify the improved network at the patch level. The experiment results show that the proposed model achieves 78.48±3.17% accuracy, 79.95±4.37% precision, 78.43±2.44% recall and 78.51±3.04% F1-score for ccRCC grading at the patch level. Furthermore, for the prediction of all patches in each patient case, the majority voting method is used to obtain the overall classification of the image level. The accuracy, precision, recall and F1-score of the proposed model at the image level are 88.89%, 89.88%, 87.65%, and 88.51%, respectively. In conclusion, the improved network with sECANet outperforms other attention mechanisms and the baseline model of ResNet50 at both patch level and image level. Therefore, the model of ccRCC ISUP grade established in this paper has relatively high diagnostic efficiency, and can even provide a certain reference for the treatment and prognosis for ccRCC patients.
In order to determine accurately International Society for Urology and Pathology (ISUP) grade of clear cell Renal Cell Carcinoma (ccRCC) and achieve subsequently better treatment and prognosis, a novel channel attention mechanism named sECANet is proposed. To obtain more useful features from the feature map, sECANet calculates the information interaction of the current channel and local channels, and calculates additionally the interaction of the current channel and remote channels. A total of 90 pathological images are collected and subsequently cut into patches. After data augmentation, 5-fold cross-validation is employed to verify the improved network at the patch level. The experiment results show that the proposed model achieves 78.48±3.17% accuracy, 79.95±4.37% precision, 78.43±2.44% recall and 78.51±3.04% F1-score for ccRCC grading at the patch level. Furthermore, for the prediction of all patches in each patient case, the majority voting method is used to obtain the overall classification of the image level. The accuracy, precision, recall and F1-score of the proposed model at the image level are 88.89%, 89.88%, 87.65%, and 88.51%, respectively. In conclusion, the improved network with sECANet outperforms other attention mechanisms and the baseline model of ResNet50 at both patch level and image level. Therefore, the model of ccRCC ISUP grade established in this paper has relatively high diagnostic efficiency, and can even provide a certain reference for the treatment and prognosis for ccRCC patients.
2022, 44(1): 149-167.
doi: 10.11999/JEIT210914
Abstract:
Residual neural Network (ResNet) is a hot topic in deep learning research, which is widely used in medical image processing. The residual neural network is reviewed in this paper from the following aspects: Firstly, the basic principles and model structure of residual neural network are explained; Secondly, the improvement mechanisms of residual neural network are summarized from three aspects of residual unit, residual connection and the entire network structure; Thirdly, the wide applications of residual neural network to medical image processing are discussed from four aspects combining DenseNet, U-Net, Inception structure and attention mechanism; Finally, the main challenges that ResNet faces in medical image processing are discussed, and the future development direction is prospected. In this paper, the latest research progress of residual neural network and its application to medical image processing are systematically sorted out, which has important reference value for the research of residual neural network.
Residual neural Network (ResNet) is a hot topic in deep learning research, which is widely used in medical image processing. The residual neural network is reviewed in this paper from the following aspects: Firstly, the basic principles and model structure of residual neural network are explained; Secondly, the improvement mechanisms of residual neural network are summarized from three aspects of residual unit, residual connection and the entire network structure; Thirdly, the wide applications of residual neural network to medical image processing are discussed from four aspects combining DenseNet, U-Net, Inception structure and attention mechanism; Finally, the main challenges that ResNet faces in medical image processing are discussed, and the future development direction is prospected. In this paper, the latest research progress of residual neural network and its application to medical image processing are systematically sorted out, which has important reference value for the research of residual neural network.
2022, 44(1): 408-414.
Abstract:
