Search

In recent years, convolutional neural networks have been widely used in various fields of computer vision and achieved remarkable results. Regularization method is an important part of convolutional neural network, which helps to avoid the overfitting phenomenon of convolutional neural network in the process of model training. There are fewer reviews on regularization methods for convolutional neural networks, and most of them lack a summary of the newly proposed regularization methods. Firstly, this paper conducted a detailed summary of the literature on regularization methods in convolutional neural networks, and classified the regularization methods into parameter regularization, data regularization, label regularization and combinatorial regularization. After that, on the public datasets such as ImageNet, it compared and analyzed the advantages and disadvantages of different regularization methods based on evaluation indexes such as top-1 accuracy and top-5 accuracy. Finally, it discussed the future research trends and work directions of regularization methods for convolutional neural network. [ABSTRACT FROM AUTHOR]

The security standards for the IoV are the cornerstone of the development of the IoV industry. To provide researchers with a comprehensive and clear understanding of the security requirements of the IoV industry, based on the three-layer architecture of the cloud-pipe device of the IoV and from the perspective of the IoV security standards, this paper statistically analyzed the IoV security standards and policy documents, compared and analyzed the focuses of the IoV security standards released by international and domestic standardization organizations, summarized the correspondence relationship between the security requirements, protection technologies and security standards of each layer, and gave the development trend of IoV security standards and construction suggestions. [ABSTRACT FROM AUTHOR]

Existing blockchain content supervision schemes adopt ex-post governance methods, lack of pre-audit, and have problems of signature invalidity and low efficiency of multi-version block verification. To address these problems, firstly, this paper designed a dynamically adjustable and accountable data audit method to realize prior audit of blockchain transaction data. Secondly, it designed a redaction-controlled digital signature scheme to solve the problem of signature invalidity caused by editing operations. Finally, it designed a blockchain data consistency verification protocol to achieve efficient verification of query results. The results of security analysis and performance testing demonstrate the security and effectiveness of the scheme. In the case of controllable supervision, the scheme still maintains high efficiency in block generation and verification, and it provides a new solution for blockchain content supervision. [ABSTRACT FROM AUTHOR]

In response to the issues of high communication overhead random primary node selection and the absence of incentive mechanisms in the three-stage process of PBFT consensus algorithm, as well as problems with, this paper introduced a new grouped consensus algorithm named DS-PBFT. Firstly, this paper optimized the consensus protocol simplified of the three-phase communication process, thereby enhancing consensus efficiency. Secondly, this paper proposed a node scoring and grouping mechanism, where nodes received scores based on their historical behavior during the consensus process and fell into consensus and candidate groups, reducing the likelihood of malicious node participation. Lastly, this paper introduced a dynamic process for selecting nodes to participate in consensus, optimized view-change protocols and garbage collection mechanisms to fundamentally reduce the number of participating nodes, resulting in enhanced consensus efficiency. Using Docker containers to simulate multiple nodes, the experimental results show that in a consortium blockchain with network stability and a significant number of trusted nodes, the proposed DS-PBFT consensus algorithm outperforms the PBFT consensus algorithm and other improved algorithms in terms of consensus latency, throughput, fault tolerance, and communication complexity. It can quickly achieve consensus and improve consensus efficiency. [ABSTRACT FROM AUTHOR]

This paper proposed a credit-score-based Byzantine fault-tolerant consensus algorithm, called CS-Raft, to address the inadequate performance of the current consensus algorithms in consortium chains. Initially, it assigned credit score attributes to all nodes and updated the credit score of each node according to their consensus behavior, activity level, time of joining the cluster and other metrics. The higher the credit score, the higher the node's trustworthiness. Secondly, it selected monitoring nodes based on their credit scores, which had the authority to inspect and participate in leader elections. The setting of monitoring nodes could effectively resist attacks from Byzantine malicious nodes. Finally, it made improvements to address the issue of vote splitting in leader elections, thereby enhancing the speed of leader election. Experimental analysis shows that compared to the PBFT algorithm, the CS-Raft algorithm effectively reduces consensus latency, increases system throughput, and accelerates leader election speed while achieving Byzantine fault tolerance. [ABSTRACT FROM AUTHOR]

To address the low utilization of raw data and insufficient feature extraction capability of multi-dimensional data in existing remaining useful life prediction methods, this paper proposed a convolutional neural network model based on feature enhancement and spatio-temporal information embedding. Firstly, it adopted a feature enhancement module to extract additional operating condition features and manual features from raw data as auxiliary features. Then, it introduced the spatio-temporal embedding module to encode the spatio-temporal information, embedding the time series information and spatial feature information into the original data. Finally, it concatenated the aforementioned features, and it employed a regression prediction module to capture the inherent relationships in the data and obtain regression prediction results. It evaluated the predictive effectiveness of the proposed model on the commonly used commercial modular aero-propulsion system simulation(C-MAPSS) dataset. The experimental results show that the root mean square error of the proposed model decreases by 8.8% on average over the four subsets compared with other mainstream deep learning methods, and it also outperforms existing state-of-the-art algorithms in prediction accuracy under multiple operating conditions and fault types. The experiments fully verify the effectiveness and accuracy of the proposed model in predicting the remaining useful life of turbofan engines. [ABSTRACT FROM AUTHOR]

Aiming at the problem that it is difficult to model DES(discrete event systems) for large-scale real systems and it is impossible to carry out effective fault diagnosis, this paper proposed a fault diagnosis method based on active learning. Firstly, the method added normal/fault labels to the acquired system event logs, divided the log set into a training set and a test set, and proposed an iterative algorithm based on abstraction technique to extract fault feature samples from the event logs in the training set. Then, it constructed the initial fault identifier from the fault feature samples, and checked the accuracy of identifier using the event logs in the test set. Simulation results show that the fault diagnosis algorithm enabled higher diagnosis accuracy under model unknown. Finally, examples illustrate the application of the fault diagnosis algorithm under system model unknown. Compared with the existing research, the proposed method can diagnose faults when the system model is unknown and the complexity of the algorithm is polynomial, which results in higher diagnostic accuracy and a wider range of applications. [ABSTRACT FROM AUTHOR]

The non maskable inter-processor interrupt(NMII) of the domestic Sunway processor must be initiated by one of the other cores. Therefore, it is difficult to apply the general lockup fault monitoring algorithm of Linux. In severe cases, it will jeopardize the data processing in critical areas. This paper designed a lockup fault monitoring and diagnosis system for Sunway architecture to solve the above problem. It used a chain structure to send NMII requests, and combined timer event and kernel thread to check the lockup time stamp, realized the soft lockup and hard lockup monitoring of single core in the system. Based on the fault tolerance mechanism, it adopted a master-slave structure to monitor the state of all cores. When the master core failed, the system implemented fault tolerance measures and migrated the master core, realized the multi-core lockup monitoring in the system. It designed a task model based on NMII, and realized the diagnostic information output of the fault cores, extended the application scenarios of NMII. The test results show that the proposed algorithm can accurately detect the lockup fault and make real-time diagnosis under both low and high fault risk, and meet the reliability and real-time requirements of lockup fault monitoring and diagnosis of Sunway platform. [ABSTRACT FROM AUTHOR]

Traditional reinforcement learning methods have problems of low efficiency, poor generalization performance, and untransferable policy models. In response to this issue, this paper proposed an inhomogeneous multitask reinforcement learning method, which improved efficiency and generalization performance by learning multiple reinforcement tasks. It constructed the morphology of agent into a graph, and the graph neural network could handle graphs with any connection pattern and size graph, which was really suitable to solve inhomogeneous tasks with different dimensions of state and action space. This breaks through the limitations that model couldn't be transferred and fully utilizes the advantages of graph neural network's natural use of graph structure to induce bias. The model had achieved efficient training and improved generalization performance, and could be quickly migrated to new tasks. The results of multi task learning experiments show that compared with previous methods, this method exhibits better performance in both multi task learning and transfer learning experiments, and exhibits more accurate knowledge transfer in transfer learning experiments. By introducing bias in the structure of the agent graph, this method has achieved higher efficiency and better migration generalization performance. [ABSTRACT FROM AUTHOR]

Autonomous vehicle can calculate the movement track of surrounding vehicles according to the track prediction algorithm, and make response to reduce driving risk, while the traditional track prediction model will produce large errors in the case of long-term series prediction. To address this issue, this paper proposed a trajectory prediction model based on the Informer algorithm, and used the publicly available dataset NGSIM to conduct experimental analysison. Firstly, it filtered the original data by using symmetric exponential moving average method(sEMA), and added a joint normalization layer to the original Informer encoder to extract features from different vehicles, reducing the motion error between different vehicles, and improving the prediction accuracy by considering the speed information of the vehicle itself and the vehicle movement information of the surrounding environment. Finally, it got the vehicle trajectory distribution at the future time through the decoder. The results show that the trajectory prediction error of the model is less than 0.5 m. Through the analysis of MAE and MSE results of trajectory prediction, when the prediction time exceeds 0.3 s, the trajectory prediction effect of Informer model is obviously better than other algorithms, which verifies the effectiveness of the model and algorithm. [ABSTRACT FROM AUTHOR]

In order to solve the problem that the existing log anomaly detection methods tend to focus only on the single feature of the quantitative relationship mode or the sequential mode, ignoring the relationship of the log time structure and the interrelation between different features, resulting in a high error detection rate and false positive rate, this paper proposed a log anomaly detection method based on the log time graph attention network. Firstly, this paper constructed a log time graph by designing a joint feature extraction module of log semantics and time structure, which effectively integrated the time structure relationship and semantic information of log. Secondly, it constructed the time relationship graph attention network, and used the graph structure to describe the time structure relationship between logs, which could adaptively learn the importance of different logs and carry out anomaly detection. Finally, it used three public datasets to verify the effectiveness of the model. Extensive experiments results indicate that the proposed method is able to effectively capture the temporal structure relationships in the logs, thereby improving the accuracy of anomaly detection. [ABSTRACT FROM AUTHOR]

To capture the complex spatial-temporal dynamics and periodic patterns in traffic flow data, and reduce the cumulative error effects caused by unexpected road conditions, this paper proposed a traffic flow prediction model based on combining PGCN and MAGRU. Firstly, the spatial-temporal data fusion module constructed periodic graphs using the property of periodic similarity in traffic flow data, and added spatial and temporal encoding information into the sequence data. Then, in the spatial-temporal feature extraction module, graph convolutional network(GCN) submodule captured spatial features from the periodic feature graphs, MAGRU submodule captured temporal features from the sequence data. Finally, the gated fusion mechanism fused the features extracted by both modules. It conducted the experiment on two real traffic flow datasets. The indicate that compared to several recent baseline models, the model achieves average reduction of 5.4%, 22.8%, 10.3% in MAE, RMSE and MAPE, exhibits an average improvement of 11.6% in R2 accuracy metric, which confirms that the model can provide more accurate predictions and reduce cumulative error effects. [ABSTRACT FROM AUTHOR]

Aiming at the problem that attributes in the existing neighborhood rough set model all have the same weight, which cannot ensure that the key attributes can be retained in attribute approximation, this paper proposed an attribute approximation algorithm based on information entropy weighting. Firstly, it adopted interclass entropy and intraclass entropy strategies to assign weights to attributes based on the principle of maximising interclass entropy and minimising intraclass entropy. Secondly, it constructed a weighted neighborhood rough set model based on weighted neighborhood relationships. Finally, it assessed the importance of attribute subsets based on dependency relationships to achieve attribute simplification. Comparison experiments with other three attribute approximation algorithms on UCI-based dataset show that the proposed algorithm can effectively remove redundancy and improve classification accuracy. [ABSTRACT FROM AUTHOR]

At present, there are only a very few works done on multi-task ordinal regression(OR). These works assume that different tasks contribute equally to the overall model. However, in practice, different tasks may have distinct contributions to the overall model. This paper proposed a novel multi-task ordinal regression method with task weight discovery method. Firstly, it presented a support-vector-machine-based multi-task OR model. By sharing the classifier parameters, the classification information could be transferred among different tasks. Secondly, considering that different tasks had different contributions to the overall model, it assigned each task a weight, which would be automatically optimized during the learning process. Finally, it adopted a heuristic framework to construct the multi-task OR model and optimized the task weights alternately. The experimental results show that the proposed method achieves 3.8% to 12.3% improvements in terms of MZE and 4.1% to 11% improvements in terms of MAE, compared to the existing multi-task OR methods. Considering the different weights of each task, and by automatically optimizing these weights, the proposed method reduces the classification error of the multi-task ordinal regression model. [ABSTRACT FROM AUTHOR]

Density peak clustering(DPC) combines the local density and relative distance of data sample points, and can cluster arbitrary shaped datasets. However, DPC algorithm has problems such as subjective selection of truncation distance, simple allocation strategy, and high time complexity. This paper proposed a density peak clustering algorithm based on grid nearest neighbor optimization(KG-DPC algorithm). Firstly, it gridded the data space to reduce the computational burden of distance between sample data points. When calculating local density, not only the density values of the grid itself were consi-dered, but also the density values of the surrounding k nearest neighbors were considered, reducing the impact of subjective selection of truncation distance on clustering results, improving clustering accuracy, and setting a grid density threshold to ensure the stability of clustering results. The experimental results show that the KG-DPC algorithm has a significant improvement in clustering accuracy compared to DBSCAN, DPC, and SDPC algorithms. Compared to DPC, SNN-DPC, and DPC-NN algorithms, the average consumption time in clustering is reduced by 38%, 44% and 44%, respectively. On the basis of ensuring the accuracy of basic clustering, the KG-DPC algorithm has specific advantages in clustering efficiency. [ABSTRACT FROM AUTHOR]

Traditional periodic pattern mining ignoreds the correlation and recency of the patterns, and therefore obtains some weakly correlated and recency patterns with limited practical value. To address the issue, this paper proposed novel periodic pattern mining methods based on recency and correlation constraints named CRPFP-BFS and CRPFP-DFS. By transforming a given database into a column-based structure CRPFP-List, CRPFP-BFS and CRPFP-DFS recursively mined its data using breadth-first and depth-first search, respectively. At the same time, CRPFP-BFS and CRPFP-DFS applied four pruning strategies based on support, period, recency and correlation to reduce the search space, thereby effectively discovering correlation recency periodic patterns. By comparing with the current state-of-the-art algorithms on dense and sparse datasets, and the experimental results show that the CRPFP-BFS and CRPFP-DFS have lower memory usage, higher operating efficiencies and good scalability. Among them, CRPFP-DFS is suitable for situations with strict memory requirements, and CRPFP-BFS performs more efficiently for long transaction sparse databases. [ABSTRACT FROM AUTHOR]

Convolution operations lead to a significant increase in the network size, which makes CNN models difficult to deploy to the embedded hardware platform, and different granularity data is not coordinated with the underlying hardware structure, which leads to low computing efficiency. Based on the reconfigurable array with the computing units supporting multiple bit widths, through software hardware cooperation and reconfigurable computing methods, this paper defined the quantization threshold using KL divergence and random integer method, proposed a strategy for finding the best basis point, designed an instruction set and a parallel mapping scheme supporting multiple bit widths to realize three distinct bit widths in data quantization. The results show the quantization scheme with 8 bit weight and feature map can compress model parameter quantity to about 50% with 2% accuracy loss. The acceleration ratios of quantifying the test images to three different bit widths reach 1.012, 1.273, and 1.556, respectively, which can shorten the execution time by up to 35.7% and reduce memory access times by 56.2%, while only bringing less than 1% relative error. This indicates that this method can achieve efficient neural network computation under three quantization bit widths, thereby implementing hardware acceleration and model compression. [ABSTRACT FROM AUTHOR]

Aiming at overcoming drawbacks of lower accuracy and being trapped easily into local optimums of the sparrow search algorithm(SSA), this paper proposed an enhanced sparrow search algorithm by integrating niche mechanism(NSSA). Firstly, NSSA adopted an adaptive niche mechanism based on producers' similarity to automatically partition the whole sparrow population into several producer-centered speciation, and to move scroungers inside the speciation towards to the producer inside the same speciation, in order to maintaining the diversity of the whole population. Secondly, NSSA applied a fuzzy logic based differential mutation operation to further mutate sparrow producers to raise the probability to jump out of local optimums. It used 12 functions selected from the CEC2017 benchmark function as the testbed to evaluate NSSA with CSSOA, ISSA, MSSSA, SHSSA, and MSWOA. Experimental results show that NSSA even can raise convergence accuracy up by 10% compared with the second ranked evaluated algorithm on most of multimodal objective test functions. This paper also conducted ablation experiments for each improved strategy proposed in NSSA. Experimental results show NSSA can achieve supreme performance by applying all 3 strategies simultaneously. At last, NSSA is applied to solve the pressure vessel design problem. Experimental results illustrate that NSSA has the ability to find the best solutions compared to other evaluated algorithms. [ABSTRACT FROM AUTHOR]

Aiming at the problems of insufficient diversity of whale optimization algorithm(WOA), inefficient and unbalanced information exchange between the two search phases, this paper proposed an improved WOA based on the armed forces collaborative warfare mechanism for community discovery. In order to solve the problem of insufficient diversity in the encircling predation stage, this paper developed a "neighbor potential" learning model to improve the global search capability and learn breadth of WOA. To solve the problem of inefficient information exchange during the two-predation phase, this paper proposed bubble net predation based on whale commanders, which could ensure effective utilization of search information. To address the imbalance between the two predation mechanisms, this paper proposed an improved learning automaton, which could guide whale populations toward promising areas. Meanwhile, because community discovery in complex networks is a discrete problem, this paper proposed a new coded discrete evolution rule based on topological properties. Finally, this paper tested the proposed algorithms on real data sets and compared them with other algorithms, and simulation experiments show that the proposed algorithm has better optimization ability than the comparison algorithm, verifying the effectiveness of the improved strategy. [ABSTRACT FROM AUTHOR]

The application of graph theory to low-order brain networks ignores the high-order dynamics of functional connections. To address these issues, this paper proposed a graph-theoretic network construction method based on high-order dynamic functional connectivity(GNC-HodFC), to analyze and classify the difference between patients with mild cognitive impairment and healthy subjects by extracting the graph theory features of higher-order FC network. Firstly, the algorithm defined graph theoretic nodes and edges that represented high-order dynamic brain network connections. Then, the algorithm used sliding window technology to extract low-order functional connection information, and put forward the stability criterion for selecting the optimal feature subset to build graph nodes. Finally, the algorithm proposed adaptive threshold strategy selection of high order dynamic functional connection information so as to build the edge of graph theory, which completed the graph construction of the higher-order dynamic brain network. The experimental results show that the average classification accuracy of GNC-HodFC is 70.5%, which is better than the other three comparison methods, and there are significant differences in the graph theory characteristics between the patient group and the healthy group. GNC-HodFC method can provide a new auxiliary means for the diagnosis of mild cognitive impairment. [ABSTRACT FROM AUTHOR]

Because of China's rapid urbanization and economic growth, the amount of waste generated from production and city life is increasing daily. Establishing effective recycling and transfer facilities is a critical problem in today's society. This paper studied the problem of finding locations for secondary waste recycling facilities, the essence of which was the NP-hard problem in combinatorial optimization. Firstly, to solve the problem, the paper made a mathematical model based on the actual situation and then studied the problem's mathematical properties to provide a proof and simplify the problem, reducing its scale and difficulty. Then, this paper designed the allocation sub-algorithm and the upper and lower bound sub-algorithms for the problem. Based on the above algorithms, it proposed a reduced-order backtracking algorithm to obtain an accurate solution while still reducing the problem size. Finally, the algorithm's principles and execution process were further elaborated through analyzing and simulating several examples. The results show that the algorithm successfully reduces the problem's difficulty by reducing the problem size. [ABSTRACT FROM AUTHOR]

Aiming at the problem that speech enhancement networks have difficulty in extracting global speech-related features and are ineffective in capturing local contextual information of speech. This paper proposed a two-branch attention and U-Net-based time-domain speech enhancement method, which used a U-Net encoder-decoder structure and took the high-dimensional time-domain features obtained from a single-channel noisy speech after one-dimensional convolution as input. Firstly, this paper designed Conformer-based residual convolution to enhance the noise reduction ability of network by utilizing residual connection. Secondly, this paper designed a two-branch attention mechanism structure, which utilized global and local attention to obtain richer contextual information in the noisy speech, and at the same time, to effectively represent the long sequence features and extract more diverse feature information. Finally, this paper constructed a weighted loss function by combining the loss function in the time domain and frequency domain to train the network and improve the performance in speech enhancement. This paper used several metrics to evaluate the quality and intelligibility of the enhanced speech, the enhanced speech perceptual evaluation of speech quality(PESQ) on the public datasets Voice Bank+DEMAND is 3.11, the short-time objective intelligibility(STOI) is 95%, the composite measure for predicting signal rating(CSIG) is 4.44, the composite measure for predicting background noise(CBAK) is 3.60, and the composite measure for predicting overall processed speech quality(COVL) is 3.81, in which the PESQ is improved by 7.6% compared to SE-Conformer, and improved by 5.1% compared to TSTNN improved by 5.1%. Experimental results show that the proposed method achieves better results in various metrics of speech denoising and meets the requirements for speech enhancement tasks. [ABSTRACT FROM AUTHOR]

In order to address the issues with the jump point search algorithm in the process of pathfinding, such as numerous path inflection points, numerous intermediate search hop points, numerous extended nodes, and lengthy pathfinding time in the process of finding jump points, this paper proposed the improved bidirectional dynamic JPS algorithm. The improved algorithm dynamically defined the target points in the forward and reverse expansion directions, dynamically defined the heuristic function, and used the dynamic constraint ellipse to restrict the expansion region of the algorithm to distinguish the expansion priorities of the inner and outer regions of the ellipse. The new cost minimum point served as the focus of the new ellipse, and the orientation and constraint region were dynamically changed as the extension from the starting point and goal point to each other. The simulation results show that the optimized bidirectional dynamic JPS algorithm has a certain performance in general maps, especially in indoor environment maps with fewer obstacles and the target point is close to the starting point. [ABSTRACT FROM AUTHOR]

The majority of current machine learning and deep learning solutions for ECG-based emotion recognition utilize fully-supervised learning methods. Several limitations of this approach are that large human-annotated datasets and computing resources are required. Furthermore, the feature representations learned by fully supervised methods tend to be task-specific with limited generalization capability. In response to these issues, this paper proposed an approach based on contrastive lear-ning for ECG-based emotion recognition, which consisted of two steps, such as pre-training and fine-tuning. The goal of pre-training was to learn representations from unlabeled EGG data through contrastive learning. Specifically, it designed two simple and efficient ECG signal augmentation methods, and used these two views to learn robust temporal representations in the time contrastive module, followed by learning discriminative feature representations in the context contrastive module. Fine-tuning used labelled data to learn emotion recognition. Experiments show that the proposed method has reached the maximum accuracy on three public ECG-based emotion recognition datasets. Additionally, the proposed method shows high efficiency under the semi-supervised settings. [ABSTRACT FROM AUTHOR]

Effectively representing features of modalities is a hot issue in cross-modal recipe retrieval. The current methods generally adopt two independent neural networks to extract the features of images and recipes respectively, achieving retrieval through cross-modal alignment. However, these methods mainly focus on the intra-modal information and ignore the intermodal interactions, resulting in the loss of some effective modality information. To address the problem, this paper proposed a cross-modal recipe retrieval method to enhance modality semantics through multimodal encoders. Firstly, it used a pre-trained model to extract initial semantic features of images and recipes and utilizing modality alignment to reduce the inter-model differences. Secondly, it employed the pairwise cross-modal attention to repeatedly reinforce the features of one modality by using features from another modality, extracted valid information. Thirdly, it used the self-attention mechanism to modal the internal features of modalities to capture rich modality-specific semantic information and potential associations. Finally, it introduced the triplet loss to minimize the distance between similar samples, achieved cross-modal retrieval learning. Experimental results on Recipe 1M dataset show that the proposed approach outperforms the current mainstream methods in terms of median ranking(MedR) and recall rate at top K(R@K), providing a powerful solution for cross-modal retrieval tasks. [ABSTRACT FROM AUTHOR]

The radar echo extrapolation method is widely used in rainfall forecasting. Addressing the issue of insufficient prediction accuracy in radar echoes, this paper proposed a deep learning model DIPredRNN based on recurrent neural networks. This model combined long-term temporal and channel information by introducing a dual attention mechanism of space and channel, improved the long-term dependence of time memory. By introducing an interactive framework of hidden states and inputs, it retained more features and improved the short-term dependence of temporal memory. This model was experimentally compared with classical models and many advanced models on the HKO-7 and Sichuan datasets. The model achieved the best results in comparing multiple indicators such as extrapolated images, MSE, SSIM, CSI-30~50 dbz. The experiment proves that the proposed DIPredRNN network improves the radar echo prediction performance and has advanced performance. [ABSTRACT FROM AUTHOR]

With the explosive growth of mobile data traffic in wireless networks, cache-enabled unmanned aerial vehicles(UAVs) are being used in mobile computing to serve as edge servers to provide on-demand services to users in the network. To bring better quality of experience to users while satisfying other resource constraints, it jointly optimized UAVs deployment, cache placement and user association to minimize content access latency for all users and provided users with content caching services of varying quality. This paper proposed a joint optimization algorithm based on alternating optimization for the scenario where multiple UAVs and ground base stations collaborate to provide caching services. The algorithm achieved a sub-optimal solution with a convergence guarantee by iteratively solving the three sub-problems obtained from the decomposition of the objective problem. Firstly, a successive convex approximation-based algorithm solved the UAV deployment sub-problem. Secondly, it used a greedy-based algorithm to solve the content caching sub-problem. Then it used a successive convex approximation algorithm based on the penalty function to solve the user access sub-problem. Finally, it repeated the above procedure to obtain a sub-optimal solution to the target problem. Several simulation experiments verified the effectiveness and feasibility of the proposed algorithm. The simulation results show that the proposed joint optimization algorithm performs better regarding average content access latency and cache hit ratio than the benchmark algorithm. [ABSTRACT FROM AUTHOR]

This paper proposed a distributed video streaming fair scheduling strategy based on federated deep reinforcement learning to address the issues of unfair user QoE and low bandwidth utilization caused by uneven video bandwidth allocation in bottleneck links. This strategy dynamically generated bandwidth allocation weights based on the client's network status and the QoE level of each video stream. The congestion control algorithm at the server side allocated bandwidth to each video stream in the bottleneck link according to the computed weights, ensuring equitable transmission of video streams in the bottleneck link. Each video terminal operated a bandwidth allocation agent, and multiple agents train periodically using federated learning to facilitate rapid convergence of the agent models. The bandwidth allocation agents synchronized their training parameters through a consensus mechanism, enabling distributed aggregation of the agent model parameters while ensuring the security of parameter sharing. Experimental results demonstrate that the proposed strategy improves QoE fairness and overall QoE efficiency by 10% and 7%, respectively, compared to the latest solutions. This indicates that the proposed strategy has potential and effectiveness in addressing the uneven allocation of video stream bandwidth and improving user experience. [ABSTRACT FROM AUTHOR]

For the channel estimation of broadband OFDM system assisted by IRS, most of the current researches are based on the full pilot setting on a single symbol, and excessive cascaded channel coefficients will lead to large pilot overhead. This paper proposed a two-stage channel estimation scheme based on block sparse in time-angle domain. Firstly, by analyzing the common sparse characteristics of the channel in the time domain and the angular domain, this paper transformed the cascaded channel matrix into a block sparse representation in the time domain-angular domain, and converted the channel estimation problem into a block sparse matrix recovery problem. Secondly, this paper proposed a two-stage sparse channel estimation scheme to sparsely recover the cascaded channel considering the transmission pilot limitation and the time-angle domain block sparse characteristics. Simulation results show that compared to the other three benchmark schemes, the proposed scheme can achieve better channel estimation performance with fewer pilots, effectively reducing the pilot overhead and achieving higher estimation accuracy. [ABSTRACT FROM AUTHOR]

At present, the mainstream solution of edge computing offloading is to model it as a multi-objective optimization problem to minimize energy consumption and delay. Different from the existing research, this paper mainly considered that tasks in different unloading areas had certain similarity in edge computing, which could be used to accelerate the convergence speed and solution effect of the algorithm. Based on this, this paper proposed an evolutionary multi-task multi-objective optimization algorithm to solve task offloading problems in different regions based on evolutionary multi-task optimization. This algorithm considered multiple independent regions to be optimized. It modeled the task offloading system model for each region as a multi-objective optimization problem, dynamically adjusting the communication level of the population by learning the user distribution in different regions and the similarity of the tasks to be processed, accelerating convergence speed, and achieving optimization for two different regions through one evolution. The experimental results show that the proposed algorithm has achieved good results in convergence speed and the uniformity of the optimal solution distribution, and can obtain the unloading deployment optimization scheme under edge computing. [ABSTRACT FROM AUTHOR]

The distributed training structure of federated learning is vulnerable to poisoning attacks. Existing methods mainly design secure aggregation algorithms for central servers to defend against poisoning attacks, but require the central server to be trusted and the number of poisoned participants to be lower than normal participants. To address the above issues, this paper proposed a poison attack defense method based on federated learning participants, which transfered the execution of defense strategies to the participants of federated learning. Firstly, each participant independently constructed a differential loss function, calculated the output of the global and local models, and conducted error analysis to obtain the weight and amount of differential loss. Secondly, it performed adaptive training based on the local trained loss function and differential loss function. Finally, this approach selected models based on the performance analysis of local and global models to prevent severely poisoned global models from interfering with normal clients. Experiments on datasets such as MNIST and FashionMNIST show that the federated learning training accuracy based on this algorithm is superior to poison attack defense methods such as DnC. Even when the proportion of poisoned participants exceeds half, normal participants can still achieve defense against poison attacks. [ABSTRACT FROM AUTHOR]

The existing RDH methods usually determine the modification model and the embedding parameters adaptively according to the given image, in order to optimize the performance. However, the existing optimization manners will lead to the high computational complexity and the unacceptable time cost. This paper proposed a modularized RDH method for JPEG images, to improve the efficiency of the adaptive optimization. The candidate embedding solutions were defined as the existing parameter sets which were verified effective under different cases. Therefore, it generated a so-called universal modification model set that was suitable for different images. Each modification model was defined as the optimal embedding points for the multiple histograms modification mechanism. During the embedding, the appropriate modification model was adaptively selected from the pre-set solution set, based on the given image. Compared with five related works on the USC-SIPI database, the peak signal-to-noise ratio(PSNR) of the proposed method can be improved 0.03~1.81 dB. In the terms of the file size increase, the results of the proposed method can reduce 12.4% and 5.1% respectively, compared with the two typical methods. The experimental results demonstrate that, the proposed method outperforms the mainstream classic methods, in the terms of the quality of the marked image and the file size preservation. Compared with the state-of-the-arts, the proposed method can achieve the similar adaptive performance, while keeping a low computational complexity. [ABSTRACT FROM AUTHOR]

With the development of urban intelligence, indoor positioning has become an important application basis for providing various location-based services. In some indoor application scenarios, the server-side needs to perform user access statistics for specific areas while ensuring the protection of user location privacy. To address this, this paper proposed a multi-level sensitive area indoor positioning algorithm based on Bloom filter and Paillier homomorphic encryption, aiming to protect user location privacy while enabling the server to judge whether a user had entered a sensitive area. The algorithm divided the indoor space based on the sensitivity level or category of areas, encrypted the data on the server-side and user-side using the Paillier algorithm, and designed an improved Bloom filter-based algorithm in the ciphertext domain to accomplish user location determination, thereby reducing the significant communication overhead and computational cost introduced by encryption operations. Experimental results on public data sets show that compared with existing spatial Bloom filter algorithms, the proposed hash array merging algorithm has a lower false positive probability with the same communication and computation overhead, and can also be extended to other applications to realize multi-class data set coding. [ABSTRACT FROM AUTHOR]

With the increasing popularity of IoT applications, the number of IoT device terminals is proliferating, the types are diverse, the levels are complex, and they are often in an uncontrollable environment, so it is critical to ensure the security and privacy of the data transmission process. This paper explored smart home services based on IoT architecture. Enabling smart home applications involved multiple parties, such as users, the cloud, ISH and smart devices, which required multi-party authentication for secure communication. This paper proposed a lattice-based multi-party authenticated key agreement protocol for smart home applications and proved that it was secure under the eCK model, and its security could be attributed to the difficulty of RLWE problem, which was resistant to quantum computing attacks. The proposed protocol is converted from an identity-based encryption scheme on lattice, does not require a public key certificate and avoids the deployment of a large PKI, achieves explicit authentication through message interaction, and can have a certain anonymity. Compared with other related post-quantum lattice-based multi-party authenticated key agreement protocol schemes, the proposed protocol is more advantageous in terms of security and execution efficiency. [ABSTRACT FROM AUTHOR]

The power Internet of Things(PIoT) is a smart service system that offers full-state awareness, efficient information processing, and convenient and flexible applications to users. However, these services also pose a risk of privacy leakage. The existing research on privacy protection of power data mainly concentrates on secure aggregation, but seldom addresses the core technology of many basic services, such as KNN query. Unlike traditional relational data, the PIoT collects flowing data of user electricity consumption, and the various power parameters exhibit dynamic correlations. Attackers can use data mining and other methods to infer future trends in data changes. Therefore, this paper proposed a privacy-preserving KNN query method. Firstly, it proposed a similarity measurement model based on bucket distance, and proved the upper and lower bounds of the error between the similarity measurement model based on bucket distance and the similarity measurement model based on Euclidean distance. Through this model, the similarity measurement could be transformed into set intersection operations. Then, it constructed a privacy-preserving function, which could generate different data privacy-preserving functions and query privacy-preserving functions for various smart terminals by substituting different parameters. Based on this, it proposed a data encoding scheme based on bucket partitioning and random number allocation. After being encrypted by the privacy-preserving function, the encoded data possessed the characteristic of ciphertext indistinguishability, and could effectively resist various attacks such as chosen plaintext attacks, data mining attacks, statistical analysis attacks, ICA attacks, and inference prediction attacks. Analysis and simulation demonstrate that the proposed secure KNN query method not only has high security but also has low overhead. [ABSTRACT FROM AUTHOR]

To solve the problem of privacy leakage and multi-task allocation in crowdsensing, this paper proposed an edge assisted crowdsensing location privacy protection(EALP) multi-task allocation mechanism. Firstly, considering the geography of tasks, this paper used improved fuzzy clustering(FCM) algorithm to cluster tasks locations, improved the clustering index, and enhanced the rationality of multi-task allocation. Secondly, to prevent collusion between the cloud and perceived users, it proposed a location privacy protection protocol in the task allocation phase. It deployed homomorphic encryption among the perceived users, cloud and edge nodes. The cloud could safely calculate the mobile distance of the perceived users without knowing their locations and the location of the task cluster center. Finally, it proposed a multi-task allocation optimization scheme based on ant colony algorithm, it balanced the interests of both platform and perceptive users by optimizing the path of execute tasks. Experiment results show that compared with similar methods, the proposed mechanism improves task completion rate while protecting location privacy, and reduces system perception costs and user mobility costs. [ABSTRACT FROM AUTHOR]

To address the issue that existing video colorization methods are complicated to guarantee both coloring quality and temporal consistency, this paper proposed a video colorization method AMVC-GAN combining attention mechanism and multi-scale feature fusion. Firstly, it proposed a GAN-based video colorization network model. It designed a multi-scale feature fusion module in the generator of GAN with a cyclic time network as the main body to obtain information of different time frequencies. Secondly, to effectively consider the relationship between adjacent frames, it used the features extracted from different time frequencies to strengthen the connection between frames as a way to enhance the temporal consistency of colorization. Finally, to obtain more helpful information, it introduced an attention module in the upsampling part, and optimally trained the results by utilizing PatchGAN to enhance the final colorization effect. Comparing with the state-of-the-art automatic video colorization methods on DAVIS and VIDEVO datasets, the results show that AMVC-GAN ranks first in multiple indicators, with better time consistency and colorization effect. Compared with other methods, AMVC-GAN can effectively reduce time flicker, while ensuring more real and natural colorization effect. [ABSTRACT FROM AUTHOR]

GAN has become one of the commonly-used image generation models. However, the discriminator of GAN is prone to the vanishing gradient problem in the training process, which leads to the instability of training. So that it is difficult to obtain the optimal GAN, and the quality of generation image is poor. To solve this problem, it designed a CNN with spectral normalization which satisfied the Lipchitz condition as the discriminator. Together with the Transformer generator, this paper proposed an image generation model, namely TCSNGAN(Transformer CSN GAN). The network structure of discriminator was simple, which could solve the problem of training instability of GAN model, and could configure the number of adjustable CSN modules according to the image resolution of data sets to achieve the optimal performance of the model. Experiments on public datasets CIFAR-10 and STL-10 show that the proposed TCSNGAN model has low complexity, and the generated image quality is good. And the experiments of fire image generation task demonstrates the effectiveness of small-sample dataset augmentation. [ABSTRACT FROM AUTHOR]

This paper proposed an unsupervised domain adaptation person re-identification base on reliability integration (UDA-RI) method aimed at alleviating the negative impact of noisy labels in the pseudo-labeling-based unsupervised domain adaptation person re-identification (UDA person ReID). This method consisted of two parts, such as progressive pseudo label refinement strategy and reliability integration strategy. The progressive pseudo label refinement strategy established a quantitative standard for measuring the uncertainty of pseudo labels and adopted gradual sampling to make the model more stable during training. The reliability integration strategy considered knowledge from different adaptation moments, allocated weights according to the reliability levels of models from different iterations, integrated the self-integrated models with different architectures, and used them as the final inference model. Experimental results show that compared with the advanced unsupervised domain adaptation person re-identification methods, the UDA-RI method achieves superior performance on Market1501, Duke- MTMC-ReID, and MSMT17 datasets. [ABSTRACT FROM AUTHOR]

Aiming at the problem of low interlayer resolution in medical tomographic images, this paper proposed an interlayer super-resolution algorithm based on deformation estimation and motion compensation. The method aimed to enhance interlayer resolution by generating interslice images. Firstly, the algorithm employed U-Net for multi-scale feature extraction and fusion of two adjacent images. To handle the complex deformation of interlayer images, it estimated slice bidirectional deformation by utilizing a warping module based on adaptive collaboration of flows. Multi-scale information fusion module performed feature aggregation on pyramid feature layers, to compensate for the motion of the generated map. Finally, it employed a post-processing network to reduce pixel outliers. The results on two CT datasets show that the PSNR of the proposed algorithm reaches 35.59 dB and 30.76 dB, the output effectively restores image details. Comparative experiments with existing methods demonstrate the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

With the continuous development of deep neural networks, significant progress has been made in the construction of cross-modal retrieval models. Cross-modal retrieval methods based on GCN have shown promising results in capturing semantic correlations in data, thus attracting increasing attention. However, most recent research focuses on incorporating correlations between labels and between samples into cross-modal representations, while the impact of correlations between label sets is neglected. In multi-label scenarios, the correlations between label sets can effectively describe semantic relationships between corresponding samples. Therefore, exploring the multi-label correlations and integrating it into cross-modal representations is important for improving the performance of cross-modal retrieval models. This paper proposed a dual semantic correlation graph convolutional networks(DSCGCN) cross-modal retrieval method. This method utilized GCN to explore the semantic correlations between labels and between multi-labels adaptively, and integrated the learned dual semantic correlations into the common representations. Additionally, it designed a multi-label similarity loss to make the similarities between the common representations more close to the semantic similarities. Experimental results on the NUS-WIDE, MIRFlickr-25K, and MS-COCO datasets demonstrate that because of multi-label correlations, DSCGCN achieves satisfactory retrieval performance. [ABSTRACT FROM AUTHOR]

This paper proposed an algorithm known as MSW-SLAM(multi-sensor information fusion SLAM based on semantic word bags) to address the issue of inaccurate LiDAR odometry position and pose calculations in the mapping of outdoor large-scale environments by mobile robots, resulting in a decrease in the accuracy of the simultaneous localization and mapping(SLAM) algorithm. This algorithm incorporated raw observation data from LiDAR using a visual inertial system and conducts joint nonlinear optimization of measurements from the inertial measurement unit(IMU), visual features, and laser point cloud features using sliding windows, the algorithm leveraged the complementary semantic word bag characteristics of vision and LiDAR for closed-loop optimization, further enhancing the global positioning and mapping accuracy of the multi-sensor fusion SLAM system. Experimental results demonstrate that, compared to traditional tightly coupled binocular vision-inertial odometry and LiDAR odometry positioning, the MSW-SLAM algorithm can effectively detect closed-loop information in trajectories and achieve high-precision global pose optimization. The point cloud map after closed-loop detection exhibits excellent resolution and global consistency. [ABSTRACT FROM AUTHOR]

Aiming at the problem that the existing video frame interpolation methods cannot effectively deal with large motion and complex motion scenes, this paper proposed a video frame interpolation method based on improved vision Transformer. This method fused the cross-scale window-based attention and the separable spatio-temporal local attention, enlarged the receptive field of attention, and aggregated multi-scale information. It jointly modeled the spatio-temporal dependencies and long-range pixel dependencies, thereby enhancing the model's ability to handle large motion scenes. The experimental results show that this model achieves PSNR values of 37.13 dB and 28.28 dB on the Vimeo90K test set and the DAVIS dataset, respectively, while the SSIM values reach 0.978 and 0.891, respectively. At the same time, visualization results show that the proposed method can produce clear and reasonable frame interpolation results for videos with large motion, complex motion and occlusion scenes. [ABSTRACT FROM AUTHOR]

For the field of autonomous driving, it is crucial to ensure accurate detection of other vehicle targets under various weather and lighting conditions. To address the limitations of the information acquired by a single sensor, this paper proposed a fusion method(AF) based on the cross-attention mechanism for fusing millimeter-wave radar and camera information at the feature level. It first spatially aligned the millimeter-wave radar and camera, projecting the aligned point cloud information into a point cloud image. Then, it expanded the point cloud image in the height and width directions to improve the match between the camera image and the point cloud image. Finally, it sent the point cloud image and the camera image into a CenterNet target detection network containing the AF structure for training and generated a spatial attention weight to enhance key features in the camera. The experimental results show that the AF structure can improve the performance of the original network in detecting targets of various sizes, especially for small targets, and has little impact on the real-time performance of the system, which makes it an ideal choice for improving the detection accuracy of vehicles in a variety of scenarios. [ABSTRACT FROM AUTHOR]

U-Net has achieved great success in the field of image segmentation. However, some of the position information is lost in the process of convolution and downsampling, model is difficult to learn global and long-range semantic interaction information and lacks the ability to integrate global and local information. To extract rich local detail and contextual information, this paper proposed an image segmentation network called MLFCNet, combining a convolutional module and a capsule encoder. Based on the U-Net, this paper introduced a capsule network module to learn target positional information and the relationships between local and global information. At the same time, the proposed attention mechanism could retain the information discarded by the network pooling layer. This paper designed a new attention mechanism so that multi-scale features could be fused, where global information was captured and background noise was suppressed. In addition, it proposed a new local feature co-occurrence algorithm to better learn the relationship between local features. The proposed method was compared with nine methods on two public datasets, mIoU improves 4.7% and Dice coefficient improves 1.7% in liver medical images compared to the second highest performing model. Experimental results on the dataset of liver and dataset of human show that under the same experimental conditions, the proposed network is superior to U-Net and other mainstream image segmentation networks. [ABSTRACT FROM AUTHOR]

In the back-end design of digital chips, it needs to consider both wire length and legalisation during global placement. Global placement represents a combinatorial optimization problem. Traditional annealing algorithms or genetic algorithms consume a significant amount of time and are susceptible to entering local optima. Current reinforcement learning solutions seldom leverage the overall visual information of the placement. Therefore, this paper proposed a reinforcement learning method that incorporated visual information to attain end-to-end global placement. During the global placement, it mapped the circuit netlist information into multiple image-level features, and utilized CNN and GCN to merge the image features with the netlist information. It employed a complete set of strategy networks and value networks to conduct comprehensive analysis and optimization of the global placement. Experiments on the ISPD2005 benchmark circuit demonstrate that the designed networks accelerate the convergence speed by approximately 7 times, reduce the placement wire length by 10% to 32%, and achieve a 0% overlap rate. This approach offers an efficient and rational solution for the global placement task of digital chips. [ABSTRACT FROM AUTHOR]

To address the problems of poor visual quality of synthesized grayscale and insufficient restoration of reconstructed color image in existing methods, this paper proposed an invertible grayscale method based on color coding and image steganography(IG-CCIS). The proposed method utilized an invertible neural network(INN) to construct an efficient color codec, and introduced dense convolutional blocks and channel attention mechanisms to further improve the performance of the network model, comprehensively reducing the loss of color information. In addition, in order to load encoded information into grayscale images and reduce image distortion caused by the embedding processed, it designed an image steganography algorithm based on exploiting modification direction(EMD). Through adaptive weight parameter selection, it could meet different embedding capacity requirements in a near-optimal manner and reduce the modification of grayscale images. Experimental tested on Kodak and McMaster datasets show that compared with existing representative reversible grayscale methods, the proposed method can generate better-quality reversible grayscale images and reconstruct more realistic color images, with better visual effects in image visualization. It also achieves better performance in terms of similarity evaluation metrics with standard reference images. [ABSTRACT FROM AUTHOR]

Copyright of Application Research of Computers / Jisuanji Yingyong Yanjiu is the property of Application Research of Computers Edition and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The concept drift and imbalance problems in complex data streams reduced the performance of classifiers.Traditional batch learning algorithms need to consider factors such as memory and runtime, but their performance is not outstanding in rapidly arriving massive data streams, and they also contain a large number of drift and class imbalance phenomena. Utilizing online ensemble algorithms to handle complex data stream problems has become an important research topic in the field of data mining. Firstly, this paper introduced and summarized the online versions of bagging, boosting, and stacking ensemble methods from the perspective of ensemble strategies, and compared the performance of different models. Secondly, this paper conducted a detailed summary and analysis of online ensemble classification algorithms for complex data streams for the first time. This paper introduced conceptual drift data stream detection and classification algorithms from two aspects: active detection and passive adaptation. This paper introduced unbalanced data streams from two aspects: data preprocessing and cost sensitivity. This paper analyzed the spatiotemporal efficiency of representative algorithms, and then it compared the performance of algorithms using the same dataset. Finally, this paper proposed the next research direction in response to the challenges in the field of online ensemble classification of complex data streams. [ABSTRACT FROM AUTHOR]

As a popular research direction in the field of artificial intelligence, Facial Expression Recognition (FER) is the most direct and effective way for computers to access human emotions. It occupies a key technical position in human-computer interaction, intelligent medical care, fatigue driving, and other R＆D topics. In order to meet the application requirements of high recognition rate, the structure of FER deep learning network becomes more and more complex, occupying a large amount of computing resources and storage space, which seriously affects the real-time requirements of the algorithm. This paper focuses on the problem of how to guarantee the accuracy of the model while effectively improving its computational speed. Firstly, it introduces the important datasets for expression recognition using lightweight networks; Secondly, it analyzes the classical lightweight network models used for facial expression recognition; Thirdly, it describes the principles, characteristics, and applicable scenarios of the main network lightweighting methods; Finally, it summarizes the problems and challenges of lightweight networks in facial expression recognition research and looks forward to the future research direction. [ABSTRACT FROM AUTHOR]