With the rapid development of multimedia,large-scale image data causes a great pressure on network bandwidth and storage.Presently,deep learning-based image compression methods still have problems such as compression artifacts in the reconstructed image and a slow training speed.To address the above problems,we propose a decoder-side enhanced image compression network under distributed strategy to reduce the reconstructed image compression artifacts and improve the training speed.On the one hand,the original information aggregation subnetwork cannot effectively utilize the output information of the hyperpriori decoder,which inevitably generates compression artifacts in the reconstructed image and negatively affects the visual effect of the reconstructed image.Therefore,we use the decoder-side enhancement module to predict the high-frequency components in the reconstructed image and reduce the compression artifacts.Subsequently,in order to further improve the nonlinear capability of the network,a feature enhancement module is introduced to further improve the reconstructed image quality.On the other hand,distributed training is introduced in this paper to solve the problem of slow training of traditional single node networks,and the training time is effectively shortened by using distributed training.However,the gradient synchronization during distributed training generates a large amount of communication overhead,so we add the gradient sparse algorithm to distributed training,and each node passes the important gradient to the master node for updating according to the probability,which further improves the training speed.Experimental results show that distributed training can accelerate training on the basis of ensuring the quality of the reconstructed image.

High-frequency(HF) communication is a communication technology which relies on the reflection of radio waves in the Ionosphere to achieve ultra-long distance transmission.HF communication also plays an irreplaceable role to guarantee the communication accessibility.However,the HF channel exhibits severe multi-path delay spread and Doppler extension,and therefore demonstrates a typical feature of time-frequency doubly-selective fading.These channel features will bring severe challenges to the reliability of HF communication.Consequently,channel equalization algorithms should be adopted at the receiver to detect the transmitted data.Conventional equalization algorithms for HF channels commonly encounter the problem of slow convergence,so that they cannot adapt to the complicated HF channel characteristics.On the other hand,the fast-converging equalization algorithms suffer from excessively high computational complexity,which fails to meet the need of engineering implementation.To address these aforementioned problems,a low-complexity orthogonal approximate message passing(OAMP) equalization algorithm is proposed for the HF doubly-selective fading channel.Specifically,we propose a LU decomposition-based iterative triangular matrix solution,and exploit the sparseness feature of the HF channel matrix.In this way,the complexity of the matrix inversion operation can be effectively reduced,and the amount of computations can be further lowered in the original OAMP equalization algorithm.Simulation results show that the proposed low-complexity OAMP algorithm demonstrates the same bit-error rate(BER) performance as the original one,while the computational complexity can be reduced by approximately 88%.

Accurate prediction of network traffic variations can help operators allocate resources and schedule in advance,thus minimizing network congestion.Existing multi-step prediction methods for network traffic struggle to capture the long-range dependencies in traffic sequences,resulting in a low accuracy in multi-step prediction tasks.In response,a novel method using time two-dimensional-variation modeling for multi-step network traffic prediction is proposed which first encodes the network traffic sequence using Gated Recurrent Units(GRUs) to accurately represent the temporal correlations of network traffic and then reconstructs the traffic based on its periodic characteristics,transforming the one-dimensional traffic sequence into two dimensions.The reconstructed traffic sequence has a compressed length and more concentrated features,enabling the model to effectively perceive its long-range dependencies.Finally,a novel convolutional neural network captures the two-dimensional features of the reconstructed traffic sequence and performs weighted fusion to produce the final prediction results.Simulation results show that compared to mainstream multi-step network traffic prediction methods,the proposed method reduces the root mean square error by at least 8.69%,mean absolute error by at least 8.96%,and the mean absolute percentage error by at least 11.73%.Experimental results demonstrate that the proposed method can effectively mine long-range dependencies in network traffic,achieving a higher accuracy in multi-step traffic prediction tasks.

In terms of maneuver model modeling,traditional maneuvering target tracking methods achieve matching between the model and the real motion of the target through adaptive interaction of the model set.When tracking non-cooperative targets,the maneuvering state changes at any time and the maneuvering forms are diverse.When the limited models in the model set cannot accurately represent its real motion,the tracking performance will degrade.This paper integrates the two level neural network of model correction and state correction into the filtering recursion process,and proposes a two-level modified maneuvering target tracking method(TLM-MTT) based on the stacked long short-term memory(SLSTM) network.The first-level model correction network perceives the maneuver of the target in real time,adjusts the model parameters,and realizes accurate modeling of the maneuver model.The second-level state correction network compensates the state estimation in real time to improve the accuracy of the filter output.The network is trained offline,and the trained network is used for online real-time tracking.Compared with traditional methods and other intelligent filtering methods,this method has a better tracking performance for high-maneuvering target tracking.

This paper investigates the design and performance optimization of transmission schemes in symbiotic radio systems empowered by intelligent reflection surfaces.In a system,the intelligent reflection surface helps the communication between the base station and the primary user.At the same time,the intelligent reflection surface acts as a backscatter device to generate secondary signals and transmit information to the secondary user by reflecting the signals transmitted by the base station.The intelligent reflection surface uses on-off keying to modulate the reflected signals by controlling the reflection units of the intelligent reflection surface.The secondary user uses energy detection to detect secondary symbols.Under the condition of ensuring the minimum transmission rate of the main transmission,the base station beamforming vector and the phase shift matrix of the intelligent reflection surface are jointly optimized to minimize the error probability of the secondary transmission.The original optimization problem is non-convex,and is transformed into two convex sub-problems by using semidefinite relaxation.The solution to the original problem is obtained by solving the two sub-problems alternatively.Finally,the computer simulation is performed to verify the theoretical analysis of the detection error probability and to evaluate the performance of the proposed scheme.The simulation results show that compared with the benchmark schemes,the proposed scheme not only improves the performance of the primary transmission system but also achieves a significantly lower bit error rate in the secondary transmission with a significantly higher bit error rate reduction rate as the transmit power,number of antennas,and number of reflective elements in the intelligent reflection surface increase.

The UAV autonomous flight obstacle avoidance technology is one of the most fundamental and critical technologies for the safe flight and application of drones,and it is also a current research hotspot in the UAV field.With the application of deep learning in computer vision and the rapid development and continuous improvement of visual sensors such as event cameras,methods about UAV autonomous flight obstacle avoidance based on visual sensors have made some progress.However,many existing research methods still face significant challenges in complex scenarios and a series of urgent problems that need to be addressed,particularly in terms of accuracy,real-time performance,and algorithm robustness.This paper first introduces the relevant concepts and difficulties of UAV obstacle avoidance.Then it categorizes obstacle avoidance algorithms based on visual sensors according to the hardware and technical approaches used into traditional obstacle avoidance methods,deep learning-based obstacle avoidance methods,event stream processing-based obstacle avoidance methods,sensor fusion-based obstacle avoidance methods,and decision-level obstacle avoidance methods based on vision.Each category of obstacle avoidance methods is discussed in detail,including their research progress and achievements,as well as an analysis of their advantages and disadvantages.Finally,the paper summarizes the existing problems in UAV obstacle avoidance algorithms and provides an outlook on future research directions.

The swift advancement of radio technology has frequently introduced radio frequency interference(RFI) at the radio observatory site,thereby contaminating the data collected from astronomical observations.When the original IQ data and the corresponding statistical features of RFI are used as inputs to a residual neural network,the temporal discontinuity of RFI hinders the convergence of loss functions,which also diminishes the recognition accuracy.To address the challenges,this paper proposes a radio frequency interference identification method based on composite time-scale features.First,the hidden information on the data is revealed through high-dimensional mapping in the time-frequency domain,whose descriptive diversity is enhanced by the fusion of both long and short time features.Second,an RFI recognition network is constructed,which consists of three parts:a deep convolutional neural network for efficient feature extraction; a path aggregation network for combining shallow graphical features with deep semantic features; a predictive output network that integrates multi-scale features for making a decision for recognition.Experimental results show that the overall recognition accuracy of the proposed method achieves 96%,representing an improvement exceeding 30% over that obtained by using the original IQ signal as the neural network input.Therefore,the method proposed in this paper effectively addresses the issue of neural networks being difficult to train and exhibiting poor performance due to the temporal discontinuity of the signals.

Using the time difference of arrival,frequency difference of arrival,and differential Doppler rate measurements from target radiation signals received by multiple receivers for joint localization is a current research hotspot as it can effectively enhance the passive location accuracy of moving targets.However,existing algorithms commonly employ a two-step weighted least squares method,which often results in significant positioning errors in low signal-to-noise ratio environments.To address this issue,a semidefinite relax method for moving targets localization is proposed.This method en-compasses three steps:first,by introducing three auxiliary variables,a pseudolinear passive location equation set is constructed; second,the localization problem is transformed into a quadratic pro-gramming problem with quadratic constraints by leveraging the relationship between the auxiliary variables and the target localization solution; finally,the semidefinite relaxation method is employed to convert the quadratic constraint optimization into a convex optimization problem,which is efficiently solved using optimization toolboxes.Simulation results demonstrate that,compared with the existing methods,the signal-to-noise ratio required by the proposed algorithm to achieve the target localization accuracy of the Cramer-Rao Lower Bound is smaller by at least 9 dB,and the proposed algorithm has a smaller localization error in lower signal-to-noise ratio environment.The effectiveness of the pro-posed algorithm is proved.

Hyperspectral image change detection has emerged as a crucial technique to identify the change of ground objects in natural scenes by incorporating abundant spectral information in hyperspectral images taken in different phases in the same area.With the thrive of deep learning,hyperspectral image change detection methods can be mainly categorized into the convolutional neural network(CNN)-based and Transformer-based method.The CNN-based methods typically adopt convolutional kernels for feature extraction,which hold the characteristics of a small receptive field and focus on local information on the image,leading to the lack of sufficient modeling of the global information.The Transformer-based methods concentrate mainly on establishing global image dependencies without taking effective local information into consideration,leading to missed or false detections in change detection tasks.To address these limitations,this paper proposes a change detection method based on multi-scale and multi-resolution information fusion.Concretely,a pyramid multi-scale high and low-frequency information extraction network is first designed to capture high-frequency details and the low-frequency content,which attach their attention on the boundary region and background region respectively at different scales of multi-temporal hyperspectral images.High-frequency information is extracted through a residual convolutional network to model local features at different scales,while low-frequency information is captured through an attention-based network to model global features.Furthermore,a dual-time-phase differential classification decision network is proposed to enhance feature extraction by adaptively learning the classification weight coefficients of each branch and generating the final weighted prediction results.The qualitative and quantitative results on three real hyperspectral datasets show that the proposed method not only showcase a superior performance on the change detection task,but also achieves a more stable and higher classification accuracy.

The existing sparse-recovery-based space-time adaptive processing(SR-STAP) method typically discretizes the angular Doppler plane into a multitude of grid points to generate a guidance dictionary.However,when these methods are employed for clutter suppression in bistatic airborne radars,they would encounter the issue of grid point mismatch,which significantly impairs the algorithm performance.In response to this problem,this paper presents an innovative approach using the atomic norm minimization(ANM) for clutter suppression in bistatic airborne radars.Unlike traditional methods,the ANM operates in the continuous domain without the need to generate a discrete grid matrix.Leveraging the positive semi-definiteness,block-Toplitz prosperity and low-rank nature of the clutter covariance matrix(CCM),the alternating direction multiplier method(ADMM) is used to iteratively solve the ANM problem,leading to the accurate estimation of the clutter subspace.Subsequently,the CCM is directly computed through eigen decomposition,improving the clutter suppression performance.Simulation results indicate that the proposed algorithm circumvents the grid-mismatch problem,achieves a more precise CCM estimation,and outperforms convolutional sparse recovery methods in terms of clutter suppression performance,particularly with fewer training samples.

Radar Non-Line-of-Sight(N-LOS) target detection is a significant issue in an urban environment,where obstacles such as buildings obstruct the line of sight between the radar and the target,thereby limiting the detection performance.Traditional methods utilize the reflection of electromagnetic waves from surfaces like walls to detect targets in blind zones.However,the attenuation of electromagnetic waves at reflective surfaces and during propagation restricts the detection range.The Reconfigurable Intelligent Surface(RIS) is a planar array controlled by low-power electronic circuits,it can manipulate electromagnetic wave reflections in the spatial domain and focus energy on desired locations by adjusting the reflection coefficients of elements.The RIS,with its low cost,ease of deployment,and low energy consumption,can be widely applied in urban environments.Inspired by the signal controlling capability of the RIS,this paper proposes the use of the RIS to aid the radar NLOS detection method,forwarding signals to bypass obstacles and reach the target.Reflection coefficients are optimized employing an alternating optimization algorithm,with the aim of improving the radar's target detection performance in N-LOS scenarios and broadening radar applications in urban environments.Simulation results demonstrate that compared with traditional methods,the RIS aided radar achieves an improved N-LOS target detection range.Additionally,increasing the number of RIS elements and positioning the RIS closer to either the radar or the target further enhances the detection performance.

While traditional unmixing methods and autoencoder-based unmixing networks have improved the unmixing performance by utilizing spatial information,they have not fully explored and leveraged spectral features.The effective integration of spectral features with spatial information could further enhance the unmixing performance.Therefore,an unmixing framework based on a Bidirectional Convolutional Long Short-Term Memory Autoencoder Network(CLAENet) with an innovative network architecture design is proposed.This framework deeply mines spatial features through convolutional layers,while convolutional long short-term memory units are used to fully explore spectral variability and the correlations between bands,effectively processing the sequential information on the spectral dimension for a more accurate and efficient analysis of hyperspectral data.To further distinguish and utilize the specificity of different spectral bands in hyperspectral data,a deep spectral partitioning method is adopted to optimize the network input.An adaptive learning mechanism is employed for refined processing of different spectral regions,enhancing the model's capability to capture complex spectral relationships within hyperspectral data and further improving unmixing performance.Comparative experiments conducted on simulated and multiple real hyperspectral datasets demonstrate that this method outperforms existing methods in terms of unmixing accuracy and model robustness.Notably,it exhibits good generalization and stability when handling complex spectral features of land cover,thus accurately estimating endmembers and abundances.

The accurate real-time traffic prediction is the fundamental technological challenge in realizing an intelligent transportation system.Current prediction methods overlook the varying degree of spatial dependence between roads when considering the spatio-temporal characteristics of traffic information,leading to a lack of differentiated design in prediction models and inaccurate predictions for individual roads.To better analyze the differences in spatial features between roads,a quantum effect traffic prediction model is designed for intuitive reasoning.This paper introduces the concept of intuitive reasoning to encode,combine,and compare road network structures,identifying highly correlated road clusters based on spatial features.The quantum annealing algorithm optimizes clustering results towards approximating global optimal solutions.Prediction models are built using the Huawei Cloud's MindSpore framework based on different clusters,focusing on the spatio-temporal characteristics within each cluster.Experiments conducted on real datasets from Los Angeles freeways in 2012 and Tokyo's 1843 freeways in 2021 are compared with various baseline models such as the History Average model,Autoregressive Integrated Moving Average model,Graph Convolutional Network,Gate Recurrent Unit,and Temporal Graph Convolutional Network.The root mean squared error performance on the two real data sets is improved by 11.32%和13.86% compared with the Temporal Graph Convolutional Network,which provides a new and effective solution to the current traffic prediction problem.

Knowledge Graphs(KG) have emerged as a pivotal trend in uncovering intricate relations between items within recommendation scenarios.While models such as the Knowledge Graph Attention Network(KGAT) focus on establishing first-order relations,their limitations become apparent when attempting to capture collaborative information inherent in high-order relations.Existing KG-based models,including the KGAT,often treat interactive behavior as a mere KG relation.They seamlessly fuse user-item bipartite graphs with knowledge graphs into a unified fusion graph,but neglect the inherent heterogeneity between graph structures.This oversight results in a weakened ability to preserve graph-specific properties.In response to these challenges,we delve into a comprehensive analysis of the latent disparities and connections between user-item interactions and relation links.We introduce a groundbreaking message propagation mechanism known as Heterogeneous Graph Information Collaborative Filtering(HGICF).This mechanism seamlessly propagates collaborative features derived from user-item behaviors and KG side information within a unified model.Unlike existing approaches,HGICF not only upholds the intrinsic attributes of inner-graph structures but also facilitates the aggregation of cross-graph information.To gain a deeper understanding of the collaborative dynamics between knowledge graphs and bipartite graphs,we introduce shared feature collaborative filtering layers.These layers are designed to adapt to various data structures and requirements by allowing different layers to be set based on the specificities of the underlying information,which leads to a flexible and adaptive model capable of capturing the nuances of diverse data sources.Extensive experiments conducted validate the superior performance of HGICF over existing state-of-the-art methods.The proposed model excels in preserving the intricate relationships within knowledge graphs and bipartite graphs,showcasing its efficacy in collaborative filtering scenarios.By addressing the limitations of current KG-based models,HGICF stands as a significant advancement in the realm of recommendation systems,offering a more robust and nuanced approach to collaborative information modeling.

As the complexity of scientific computation increases,workflows have become an essential model for automating scientific computations.Workflow as a Service(WaaS) platforms rent virtual machines from Infrastructure as a Service(IaaS) providers to offer users the service of running scientific workflow computations.However,current researches on workflow scheduling in WaaS platforms do not consider the potential for virtual machine downtime to lead to task failures and the delays in virtual machine provisioning.To address this issue,this paper proposes a multi-workflow fault-tolerant scheduling strategy for WaaS platforms.First,considering that WaaS platforms do not schedule hardware resources but operate at the level of virtual machines and containers,we establish a workflow scheduling model suitable for WaaS platforms,taking into account the impact of virtual machine provisioning delays on scheduling.Second,we propose a multi-workflow fault-tolerant scheduling strategy for WaaS platforms,which includes preprocessing,fault-tolerance selection method,task scheduling,and resource adjustment.This involves designing an improved deadline division algorithm for determining the scheduling order,creating a fault-tolerance selection algorithm that combines replication and resubmission,considering task attributes and virtual machine provisioning delays for virtual machine selection and task allocation,and designing a resource adjustment algorithm for avoiding the waiting time for the provisioning delay of virtual machines or containers by deploying resources in advance for the upcoming tasks.Finally,by comparing the proposed scheduling strategy under different virtual machine downtime probabilities,workloads,and deadlines with other algorithms,we demonstrate the effectiveness of the proposed fault-tolerant scheduling strategy for WaaS platforms.

In order to further study the application of deep learning in cryptographic security analysis,neural networks are used for differential analysis of lightweight block cryptography.The following four research results are obtained.First,a neural network differential distinguisher is constructed by using a deep residual network with an attention mechanism,and applied to three types of lightweight block ciphers:SIMON,SIMECK and SPECK.The results show that the effective distinguisher of SIMON32/64 and SIMECK32/64 can reach up to 11 rounds,and the accuracy is 0.5172 and 0.5164,respectively.The SPECK32/64 has an effective distinguisher of up to 8 rounds with an accuracy of 0.5868.Second,the influence of different input differences on the accuracy of the neural network differential distinguisher is explored.For SIMON,SIMECK and SPECK ciphers,the accuracy of the neural network differential distinguisher corresponding to different input differences is obtained by using the fast training of neural networks.The results show that the input difference with a low Hamming weight and high probability can improve the accuracy of the neural network differential distinguisher.At the same time,the suitable input difference for the SIMON32/64,SIMECK32/64 and SPECK32/64 neural network differential distinguisher is found to be 0x0000/0040,0x0000/0001 and 0x0040/0000,respectively.Third,the influence of the input data format containing different information on the accuracy of the neural network differential distinguisher is explored.Changing the amount of information contained in the input data according to the characteristics of the cryptographic algorithm and retraining the corresponding neural network differential distinguisher.The results show that,compared to a neural network differential distinguisher that only includes ciphertext pair information,those that incorporate both ciphertext pair information and differential information from the penultimate round achieve a higher accuracy.Fourth,on the basis of the above research,the last wheel key recovery attack is carried out on 11 rounds of SIMON32/64.When 2⁹ plaintext-ciphertext pairs are selected,the attack success rate in 100 attacks can reach 100%.