To address the issues of inconvenient operation,low efficiency,and difficulty in managing recognition data during the manual identification of substation layout drawings,this paper proposes a morphology-based large-size drawing segmentation method and a text-image multimodal fusion drawing symbol detection method.Combined with post-processing methods for symbol detection,this forms a detectable and adaptable approach to large-size layout drawing symbol detection that can be generalized to other fields.The text-image multimodal fusion drawing symbol detection model is improved upon the open-set object detection model YOLO-World,by introducing the CTCM,SOFEM,and CJFFM.These enhancements significantly improve the model's performance in symbol recognition.Using the proposed methods,the detection of symbols in actual high-speed railway traction substation general layout drawings dataset is achieved.Compared to the original model,the proposed improved model,while maintaining a similar level of complexity,reaches an average precision of 97.5% for symbol recognition,with mAP@50:95 and mAP@90 increasing by 1.1% and 3.0%,respectively.
In recent years,the technologies of LLMs have been rapidly developed,with their applications across various industries experiencing vigorous growth.From natural language processing to intelligent recommendations,and from information retrieval to automated writing,LLMs are becoming indispensable tools in many fields.However,with the diversification of application scena-rios and the increase in demands,the efficiency of LLM inference is becoming increasingly prominent.In practical applications,ra-pid and accurate inference capabilities are crucial for responding to user queries,handling large-scale data,and making real-time decisions.To address this challenge,academia has undertaken extensive research and exploration to enhance the inference efficiency of LLMs.This paper comprehensively surveys the literature on efficient LLM inference in practical application scenarios.Firstly,it introduces the principles of LLMs and analyzes how to improve LLM inference efficiency in practical application scenarios.Secondly,it proposes a taxonomy tailored for real-world applications,which consists of three main levels:algorithm optimization,parameter optimization,and system optimization.This survey summarizes and categorizes related work about LLMs.Finally,it discusses potential future research directions.
Event cameras based on neuromorphic spike signals can provide information about illumination changes,compensating for the performance degradation of traditional RGB cameras in object detection under adverse environments.However,existing methods fusing event cameras with conventional cameras suffer from large model parameters and non-end-to-end training approaches,which restrict the effectiveness of modality fusion.To address this,this paper proposes a lightweight end-to-end object detection framework that integrates event and RGB information through multi-granularity fusion of multi-scale features across different network levels.By implementing lightweight fusion modules with reparameterized convolutions and enabling end-to-end training,the proposed framework enhances the model’s capability to extract complementary information from both modalities,overcoming challenging conditions in autonomous driving.Evaluated on the large-scale PKU-SOD dataset containing vehicular visual data under low-light,high-speed motion blur,and normal illumination scenarios,the proposed method significantly reduces model parameters compared to state-of-the-art multimodal approaches while improving detection accuracy and inference speed,demonstrating superior performance over existing methods.
In social media,people’s comments usually describe a certain sentiment region in the corresponding image,and there is correspondence information between image and text.Most previous multimodal sentiment analysis methods only explore the interactions between images and text from a single perspective,capturing the correspondence between image regions and text words,leading to results that are not optimal.In addition,data on social media is strongly personal and subjective,and the sentiment in the data is multidimensional and complex,which leads to the emergence of data with weak image and text sentiment consistency.To address the above two problems,a multimodal sentiment analysis model with interactive fusion of two perspectives under cross-modal inconsistency perception is proposed.On the one hand,cross-modal interaction of graphic and textual features from both global and local perspectives provides a more comprehensive and accurate sentiment analysis,which improves the perfor-mance and application of the model.On the other hand,the inconsistency scores of the graphical features are calculated to representthe degree of graphical inconsistency,as a way to dynamically regulate the weights of the unimodal and multimodal representations in the final sentiment features,thus improving the robustness of the model.Extensive experiments are conducted on two public datasets,MVSA-Single and MVSA-Multiple,and the results demonstrate the validity and superiority of the proposed multimodal sentiment analysis model compared to the existing baseline models,with F1 values increasing by 0.59 persentage points and 0.39 persentage points,respectively.
Screen-shooting resilient watermarking algorithms are of significant importance in fields such as copyright protection and traceability.Existing screen-shooting resilient watermarking algorithms mostly focus on natural images,neglecting research on document images.Document carriers inherently contain less redundant information,making it challenging to balance robustness and imperceptibility of the watermark.To address this issue,a screen-shooting resilient watermarking method for document image based on attention mechanism is proposed.To enhance the imperceptibility of the watermark,an attention feature fusion module is introduced in the encoder network to adaptively aggregate shallow and deep features.To improve the robustness of the algorithm for extraction,an adaptive channel-spatial attention module is designed in the decoder network to emphasize features that are particularly important in both channel and spatial dimensions.Additionally,a Moiré distortion layer is designed during screen-shoo-ting noise simulation to enhance the algorithm’s robustness against real Moiré distortions.Experimental results demonstrate that the proposed method achieves an average PSNR of 33.4 dB,SSIM of 0.988 5,RMSE of 5.48,and an average extraction accuracy of 99.49% in various screen-shooting scenarios.In terms of image quality and watermark robustness,the proposed method outperforms existing similar methods.
Time series data is widely present in fields such as finance,healthcare,industry,and transportation.Time Series Ano-maly Detection(TSAD) is crucial for ensuring system stability and safety.Most current time series anomaly detection methods are unsupervised due to the difficulty in collecting anomaly samples.However,these methods commonly face the problem of over-generalization,where the model can not only reconstruct normal samples,but also effectively reconstruct anomaly samples,leading to poor anomaly detection performance.Therefore,this paper proposes a time series anomaly detection method based on Kolmo-gorov-Arnold representation theory,called TS-KAN.TS-KAN leverages its parameter efficiency and local plasticity to better fit normal samples and alleviate the overgeneralization problem.Additionally,this paper introduces a local feature enhancement layer,namely Local-KAN,to enhance the representation of temporal features and improve contextual anomaly detection capability.Experiments on five mainstream time series anomaly detection datasets demonstrate that TS-KAN significantly outperforms existing methods in anomaly detection capability.
Tactical confrontation simulation experiments are the core means of operational analysis,simulation training,and equipment activities based on simulation,and their levels of intelligence and automation directly impact the effectiveness of experiments and the generation of combat capabilities.To address the low efficiency issues in experimental design,model construction,scenario control,and human-computer interaction in traditional simulation experiments,a system architecture for empowering tactical confrontation simulation experiments with large language models is proposed,referencing the MCP protocol.This architecture consists of five layers:the foundation layer,tool resource layer,AI agent layer,empowerment path layer,and application layer.The five-layer architecture is guided top-down and integrated bottom-up layer by layer,enabling the coupling and aggregation of large and small models with data resources and traditional small models,and empowering various military activities based on simulation.Based on this,the specific paths of large model empowerment in tactical confrontation simulation are discussed in detail:large model empowerment in simulation experiment design,large model empowerment in decision-making model construction,and large model empowerment in scenario control.Finally,the challenges and countermeasures are analyzed.
Most of the existing multi-view clustering algorithms only rely on the low-order similarity information between views,fail to capture the high-order structural features in the data effectively,and pay insufficient attention to the diversity features of the multi-view data,resulting in the accuracy and robustness of the clustering results.To solve these problems,JSEC,a joint spectral embedding clustering algorithm based on multi-view diversity learning,is proposed.Through view diversity learning,multiple features between data are preserved,so as to effectively remove the noise in the view.Then,a method of mining higher-order information of views is proposed to make the diversity features of views as close as possible to the hybrid similarity graph,so as to realize efficient integration of information of different views,and realize the diversity and complementary integration of views.Finally,the diversity feature matrix of the view is fused into the joint spectral embedding matrix in the spectral embedding module,and the graph clustering is realized by spectral clustering.In addition,an alternate iteration method is designed to optimize the objective function.In comparison with the latest multi-view clustering algorithms,JSEC algorithm shows superior performance on 3 indicators of 5 medium and small scale real datasets,and also on 2 large scale datasets.Compared with the suboptimal algorithm,ARI index has an improvement of 1.27% and 2.57% ondatasets of different scales.The superiority of the algorithm is proved theoretically and experimentally.
The research on energy-efficient scheduling of synchronous tasks in multi-core real-time systems mainly focuses on homogeneous multi-core processor platforms.The architecture of heterogeneous multi-core processors can more effectively exert the system performance.If the existing research is directly applied to heterogeneous multi-core systems,guaranteeing schedulabi-lity may lead to higher energy consumption.Using DVFS technology,the energy-efficient scheduling problem based on task synchronization in heterogeneous multi-core real-time systems is studied,and an algorithm named SA-LESF(Synchronization Aware-Largest Energy Saved First) is proposed.The algorithm iteratively optimizes the speed configuration of all tasks until all tasks reach their maximum energy-saving speed configuration.In addition,the SA-LESF-DR(Synchronization Aware-Largest Energy Saved First with Dynamic Reclamation) based on reusing dynamic slack time is further proposed.While ensuring that real-time tasks can be scheduled,the algorithm implements corresponding reusing strategies to further reduce system energy consumption.The simulation results show that the SA-LESF and SA-LESF-DR algorithms have advantages in energy consumption perfor-mance,under the same task set,it can save up to 30% more energy compared to other algorithms.
In recent years,prompt learning has been widely applied in the field of natural language processing.According to research,argument roles are highly semantically related to topics in text,and existing prompt tuning methods overlook entity information and interactions between arguments.Therefore,this paper proposes a bidirectional prompt tuning event argument extraction model(TEPEAE) that integrates topic and entity embeddings.Firstly,topic features are extracted using a topic model and embedded into a topic representation.Secondly,prompt templates are constructed based on trigger words,arguments,and entity information,incorporating topic embeddings into the template.Thirdly,masked language model(MLM) is utilized to predict the role label for each entity.Finally,labels are mapped from the label word space to the argument role space.Experiments on ACE2005-EN and ERE-EN datasets show that TEPEAE outperforms baseline models and achieves 79.53% and 78.60% in terms of F1,respectively,which demonstrates the effectiveness of TEPEAE.Moreover,it continues to demonstrate exceptional performance in low-resource scenarios,further proving its enhanced robustness.
Graph data,as a data type with strong expressive power,is difficult to model efficiently due to its complex structure.How to effectively capture its intrinsic information has become a challenging problem.Graph embedding methods have received increasing attention by mapping high-dimensional sparse graphs into low-dimensional dense feature vectors,while preserving the structural information of graphs.However,the existing reviews do not summarize the graph embedding methods comprehensively enough,especially paying less attention to complex graph embedding,which leads to the failure to systematically sort out the current status of research on graph embedding in dealing with diverse graph data.Therefore,this paper presents a systematic review of graph embedding learning methods from simple to complex graphs.Firstly,it gives the common definitions of various types of graphs and graph embedding.Secondly,it systematically summarizes the embedding methods on simple graphs,including shallow and deep embedding methods.Then,it summarizes the embedding methods on complex graphs according to the types of graphs,focusing on the application of deep embedding techniques in complex graph structures such as dynamic graphs,heterogeneous graphs,multiplex graphs,and hypergraphs,to fill the gaps in the existing literature that is insufficiently researched on complex graph structures.Finally,it discusses the practical application scenarios of graph embedding techniques,and looks forward to the future development directions.
Deep neural networks(DNNs)have long been threatened by adversarial attacks,particularly sparse attacks in black-box attacks.These attacks rely on the target model's output to guide the generation of adversarial examples and typically deceive image classifiers by perturbing only a few pixels.However,existing sparse attack methods suffer from inefficiencies due to the use of fixed step-size strategies and poor initialization approaches,which fail to fully exploit perturbations.To address these issues,SSA is proposed.Unlike other methods that use fixed step sizes,SSA adapts the step size based on historical search information,thus accelerating the discovery of adversarial examples.Additionally,recognizing that sparse attacks in black-box settings tend to perturb high-importance pixels,SSA uses an initialization strategy based on the CAM,interpretability method,to quickly identify and initialize populations of high-importance pixels.Finally,to confine perturbations within critical sections and maximize their ef-fectiveness during the search process,SSA adopts a section search strategy to reduce the search space.Experimental results de-monstrate that SSA outperforms the SOTA(State-of-the-Art)methods,in attacking traditional convolutional networks and Vi-sion Transformer(ViT)models.Specifically,SSA achieves a 2％～8％improvement in attack success rates and approximately a 30％enhancement in efficiency.
To address the challenges of highly integration with the environment and the high demand for real-time performance,this paper proposes a camouflaged object detection model for aerial images using bidirectional cross-attention cross-domain fusion.Firstly,a feature extraction network with two branches is constructed to extract features from both RGB and frequency domain.Simultaneously,frequency features and RGB features are crossly fused at multiple scales using bidirectional cross-attention fusion modules,effectively improving the network’s representational capacity.Experimental results show that the proposed model achieves a better balance between target recognition accuracy and real-time performance,compared to other representative models.
Deep neural networks are vulnerable to adversarial example attacks.Existing transfer-based attack optimization methods commonly employ fixed constraint upper bounds to represent imperceptibility intensity,focusing primarily on improving attack success rates.However,such approaches overlook inter-sample sensitivity variations,resulting in suboptimal imperceptibi-lity(measured by Fréchet Inception Distance,FID).Inspired by adaptive gradient methods,this paper proposes an adversarial at-tack optimization method with adaptive constraint upper bounds,aiming to enhance imperceptibility.Firstly,a sensitivity metric based on gradient magnitudes is established to quantify sensitivity differences across samples.Building on this,adaptive constraint upper bounds are determined to enable differentiated perturbation handling-applying low-intensity perturbations to sensitive samples and high-intensity perturbations to non-sensitive ones.Furthermore,by replacing the projection operator and step size,the adaptive constraintmechanism is seamlessly integrated into existing attack methods.Experiments on the ImageNet-Compatible dataset demonstrate that,under equivalent black-box attack success rates,the proposed method reduces FID by 2.68％～3.49％compared to traditional fixed-constraint methods.Additionally,the MI-LA attack algorithm based on this approach achieves 6.32％～26.35％lower FID than five state-of-the-art adversarial attack methods.
Multi-agent decision-making offers extensive engineering applications,particularly in the cooperative control tasks.Po-licy gradient-based reinforcement learning methods,which directly model policy distributions,are more conducive to exploring diverse strategies in complex reward scenarios.These methods also demonstrate consistently high empirical efficiency across both discrete and continuous action spaces.Although parameter-sharing mechanisms are widely adopted in policy gradient frameworks to improve convergence efficiency for collaborative tasks,the lack of attention to action semantic modeling introduces critical limitations,especially in mitigating action homogenization among agents.To solve this issue,this paper proposes CSF method from a graph-based modeling perspective.The CSF framework employs a graph autoencoder to learn correlation-aware semantic embeddings within the action space,subsequently achieving information fusion through dynamic integration of agent-specific beha-vioral features with semantic embeddings.This fusion mechanism aggregates collaborative behavioral information into agent-specific latent representations,enabling interdependent policy space exploration across agents.Comprehensive experiments conducted on diverse complex task scenarios within the StarCraft and Google Research Football environments demonstrate that CSF achieves superior performance over state-of-the-art algorithms,thus validating its effectiveness in facilitating inter-agent collaboration.
Attribute grouping is one of the effective steps in high-dimensional outlier detection,but the current ensemble strategies in attribute grouping-based outlier detection only take into account the local outlier information within each attribute group,and ignore the global outlier information of all attribute groups,which can lead to a biased ensemble of attribute group outlier information.This paper proposes an attribute grouping outlier detection approach based on Isolated Forest ensemble strategy by using the local and global outlier information of attribute groups.Firstly,attributes are automatically divided into several attribute groups based on the local and global correlation among attributes,and the outlier information of data objects is obtained in each attribute group.Secondly,from the perspective of attribute grouping,the ensemble bias of the current outlier information ensemble strategy is theoretically analyzed,and the ensemble deviation coefficient are defined as the evaluation index of the outlier information ensemble strategy.Then an attribute grouping-based isolation forest ensemble strategy for categorical outlier detection is proposed,this strategy effectively depicts the local and global outlier information of attribute groups and lowers the ensemble bias of attribute group outlier detection.In the end,experimental results on the UCI validate that the ensemble strategy effectively alleviates the ensemble bias and improves the outlier detection performance.Importantly,compared with the competing methods,the algorithm bolsters the AUC index and the detection efficiency by averages of 7.83% and 48.43%.
Large language model-based agent operating systems(Agent OS),as core platforms for integrating large models,tool resources,and multi-agent collaboration,are gradually becoming a key research direction for advancing general artificial intelligence.This paper systematically reviews the research progress in the field of Agent OS.It begins by discussing foundational theories,reviewing the evolution of various large language models,and progress in agent technology and traditional operating systems.This paper then elaborates on how their hierarchical architectures and modular designs achieve resource management and intelligent scheduling,focusing on typical architectures such as AIOS.Furthermore,it clarifies existing technical bottlenecks in scalability,context integration,and security within current systems.It also proposes future directions,including the use of lightweight designs,self-supervised learning mechanisms,and dynamic scheduling algorithms to optimize multi-agent cooperation efficiency.The main contributions of this paper are integrating fragmented research to provide a clearer technical framework,and highlighting the current limitations of Agent OS in covering emerging applications and industry-specific customizations.Future work should focus on enhancing the capability of cross-domain Agent OS for self-evolution and accelerating their implementation across diverse fields.
Flexible job shop scheduling problem(FJSP),as a variant of the job shop scheduling problem,has become an important research topic in the intelligent transformation of modern manufacturing industry due to the wide applicability.In recent years,deep reinforcement learning(DRL) has been applied to solve flexible job shop scheduling problems.However,the characteristic that operations can be assigned to multiple compatible machines with different processing times brings additional complexity to decision making and state representation.This paper proposes an end-to-end deep reinforcement learning framework based on an improved attention mechanism and proximal policy optimization algorithm to solve the FJSP.Considering the characteristics of heterogeneous disjunction graph structure,it designs a double-layer attention network based on hierarchical attention,including node-level attention layers and type-level attention layers,to fully extract the complex information between operations and machines to support high-quality scheduling decisions.Experimental results on synthetic and public datasets show that the proposed method outperforms traditional priority dispatching rules and currently state-of-the-art DRL methods in both of performance and generalization ability while maintaining high efficiency.
The standard SDP architecture employs identity-based authentication and authorization strategies to monitor and audit access activities in real time.However,users must fully disclose their identity information to obtain access,potentially exposing sensitive data unrelated to the requested service and introducing privacy risks.To address challenges such as ineffective user privacy protection and vulnerability of access records to malicious linkage in the current SDP architecture,this paper proposes an anonymous authentication scheme based on verifiable credentials(VCs) for SDP.The scheme constructs a VC verification algorithm using bilinear pairing and CL-signature,integrating the VC system with the standard SDP architecture to enable anonymous user access without altering the original single-packet authorization and TLS secure connection authentication model.Theoretical analysis demonstrates that the proposed scheme resists common network attacks,including knock amplification and identity impersonation.Experimental results show that it achieves shorter authentication latency in multi-node network environments.
The discrete logarithm problem is a significant challenge in number theory,and due to the difficulty of solving it,classical computers lack efficient algorithms for this task.As a result,the discrete logarithm problem is widely used in public key cryptosystems,and if it were cracked,it would pose a direct threat to the security of these systems.However,with the advent of quantum computing,researchers have begun exploring quantum computers as a potential solution for the discrete logarithm problem.Currently,most quantum algorithms for solving the discrete logarithm problem are based on Shor’s algorithm.However,due to Shor’s inherent limitations,these algorithms often face issues such as large quantum circuit depth,high qubit usage and complex post-processing steps.This makes it difficult for Shor’s algorithm to be implemented on NISQ computers.To address these issues,this paper proposes a novel approach by introducing a variational quantum algorithm for solving the discrete logarithm pro-blem.This algorithm leverages the parallelism of quantum computing to compute the modular exponentiation of parameterized quantum states.It also designs a marked solution circuit that maps valid solutions of the discrete logarithm problem onto auxiliary qubits.Then,a classical optimizer is used to iteratively adjust the parameters within the parameterized quantum circuit,continuously reducing the value of a designed loss function.Finally,the optimized parameters from the classical optimizer are fed into the measurement circuit,where the solution to the discrete logarithm problem can be obtained with high probability.Compared to Shor’s algorithm,the proposed method significantly reduces the required number of qubits and nearly halves the quantum circuit depth.Furthermore,this paper provides a detailed design of the quantum circuit and verifies the correctness of the proposed algorithm using the Qiskit package in Python.
