最新刊期
卷 25 ， 期 7 ， 2026
- 摘要：Addressing the challenge of high-quality scheduling of dependent tasks in multi-agent collaborative software development， which is hindered by unknown， heterogeneous， and noisy agent capabilities， we construct a five-tuple formalized model S =＜T， A， F， M， Π＞ from a software engineering perspective， encompassing task decomposition， capability mapping， and feedback optimization. We propose an iterative scheduling algorithm， BayesSchedule， driven by Bayesian capability learning as its core and with UCB exploration and matching threshold as an optional enhancement module. The experiments are based on the Pegasus real-world workflow benchmark and the Bernoulli-Gaussian failure model. Five sets of simulation experiments are set up， covering scale， noise， and structural conditions， and are compared with three online learning baseline methods： UCB1， Thompson Sampling， and QL-HEFT. The experimental results show that after 100 iterations， the capability estimation error of BayesSchedule drops to 0.01. The total scheduling cost is basically on par with the theoretical optimal benchmark Oracle. The total task duration （makespan） is reduced by 10.0% compared to HEFT， and the total cost is reduced by 30.2%， 28.3%， and 67.9% compared to the three online learning baselines， respectively. The research results validate the effectiveness of the feedback mechanism F and scheduling strategy Π in capability uncertainty scenarios， and provide a formalized basis for multi-agent collaborative software development to evolve from "tool assistance" to "capability orchestration".关键词：multi-agent system;software engineering;collaborative development;formal modeling;Bayesian scheduling;simulation experiment1|0|0更新时间：2026-08-03
-
An Interpretable Primary Headache Auxiliary Diagnostic Model Driven by Knowledge and Data Fusion AI导读
摘要：Primary headache is a common chronic disease in neurology， and the diagnostic process relies on symptom feature recognition and clinical experience. The International Classification of Headache Diseases， 3rd edition ICHD-3， provides criteria for the diagnosis of headaches， but many entries are qualitative descriptions with unclear boundaries， making it difficult to model directly. Traditional machine learning is difficult to simultaneously process clinical rules and complex features； Although deep learning can improve classification performance， its interpretability is insufficient and its clinical use is limited. To this end， a knowledge and data fusion driven interpretable primary headache auxiliary diagnosis model is proposed. Firstly， the model consists of two parts： hierarchical fuzzy logic and KAN network： interpretable mapping of ICHD-3 standard using fuzzy rules； Efficiently modeling the nonlinear relationship between multidimensional symptom features through the function approximation ability of KAN network. Secondly， the SHAP method is introduced to analyze the importance of model features and optimize the learning process of diagnosing key variables. Finally， embedding medical rule knowledge into the deep learning training process enables collaborative optimization and dynamic weight adjustment of diagnostic rules and data-driven approaches. The experiment shows that the proposed fusion model outperforms the comparison method in terms of accuracy （0.958 4）， sensitivity （0.951 3）， specificity （0.961 8）， and Kappa value （0.888 1）， demonstrating superior diagnostic performance. The visualization analysis of the model structure and reasoning process shows that the model has excellent clinical interpretability and strong practical value and promotion prospects.关键词：primary headache;explainable artificial intelligence;hierarchical fuzzy logic;Kolmogorov-Arnold network;knowledge-data fusion1|0|0更新时间：2026-08-03 - 摘要：In view of the challenges of multimodal brain network feature fusion in schizophrenia， such as strong modal heterogeneity， complex structure function coupling relationship， and difficulty in dynamic modeling of multimodal feature contributions， a multimodal dynamic fusion graph network model， MDF-GNN， is proposed. This model realizes multimodal information fusion through a triple mechanism. Firstly， independent and shared convolutional channels are designed and combined with adaptive sensing pooling （ASAP） to extract the sub network characteristics of functional connectivity （FC） and structural connectivity （SC）； Secondly， the dual attention mechanism is used to dynamically balance the feature contribution weight in the shared channel and the feature contribution weight in the multi-channel； Finally， the class discrimination is enhanced by constructing a joint representation space of contrastive learning constraints. The model achieved a classification accuracy of 93.55% on the self collected schizophrenia data set， and achieved a performance increase of 16.13% compared with the best single-mode baseline model GCN （accuracy of 77.42%）. The results of visual analysis showed that the combined contribution of structural （cortical thickness） and functional （local consistency） characteristics in the patient group accounted for 65.3%， suggesting the synergistic effect of gray matter atrophy and functional compensation imbalance； The results of multi-channel weight analysis showed that patients were more dependent on SC （the weight increased by 17% compared with the healthy group）， while the healthy group was dependent on FC （62.9%）， which confirmed the neurodevelopmental mechanism of the disease； The difference in the weight of shared channels between groups （Δ+4%） suggested that there was structural functional decoupling in patients. The above results provide new ideas for the development and precise diagnosis and treatment of multimodal biomarkers.关键词：magnetic resonance imaging;graph convolutional network;multimodal fusion;comparative learning;attention mechanism1|0|0更新时间：2026-08-03
- 摘要：Compared to centralized data centers， Geographically Distributed Data Centers （GDDC） can significantly reduce Wide Area Network （WAN） transmission latency and enhance system fault tolerance through proximity-based resource deployment and load balancing mechanisms. However， their vast scale， heterogeneous resources， and the differences in inter-regional electricity prices and network costs make it a key challenge for GDDC to optimize costs while ensuring quality of service. In response， this paper proposes a deadline-aware and cost-driven heuristic scheduling algorithm， Cost-effective Aware Dynamic Job Ranking （CADR）. The core mechanism of CADR is to dynamically adjust the priority of tasks in list scheduling by scaling the reserved time and expected cost， thereby prioritizing the selection of low-cost resources while meeting task deadlines. Experimental results show that compared with mainstream algorithms such as PRESTO， HEFT， PEFT， and MSL， CADR can reduce the average cost by up to 47% with only a slight increase in execution time， significantly improving the cost-effectiveness of Geographically Distributed Data Centers.关键词：cloud computing;geo-distributed data centers;cost minimization;task re-ranking1|0|0更新时间：2026-08-03
- 摘要：In view of the inherent distributed characteristics of federated learning， its data is stored in multiple participants and each participant independently trains the model. The lack of centralized global data control and unified model verification mechanism makes it highly vulnerable to poisoning attacks. Existing research mainly focuses on intercepting updates from malicious clients under the single global model architecture， and once the interception fails， it will directly lead to poisoning of the global model. Therefore， a federated learning grouping method dFL is proposed to detect abnormal updates on the client side. First of all， the client is divided into multiple groups by grouping idea， and each group trains a group model. Even if there are malicious clients， the scope of influence is limited to the group model； Secondly， in the intra group training， the abnormal update detection mechanism is used to combine the client reputation value for weight aggregation to further reduce the impact of malicious clients on the group model； Finally， the final result of the global model is voted by the group model. On MNIST-0.5 and CIFAR10 datasets， dFL shows stronger stability than the control algorithm.关键词：federated learning;federated group learning;weight aggregation;anomaly detection1|0|0更新时间：2026-08-03
Artificial Intelligence
- 摘要：As a pivotal component within the equipment testing and evaluation system， military software capability requirement compliance assessment directly influences the effectiveness of combat efficacy generation and verification. Addressing current challenges in evaluation methodologies， including ambiguous expert-experience-based indicators， insufficient adaptability between static weighting models and dynamic multi-task coupling effects， as well as inflated weights of cross-dependent indicators， this study draws inspiration from software scenario testing. We propose a dynamic scenario flow graph modeling approach to deconstruct typical operational task sequences， coupled with Monte Carlo simulation to establish a dynamic quantification model for scenario criticality weighting. Building upon this foundation， to resolve the persistent issue of cumulative deviation caused by redundant accumulation of shared indicator weights in traditional hierarchical evaluation， we design a mutual-feedback optimization mechanism based on Contribution Ratio Coefficient （CRC）. This leads to the development of a hierarchical TOPSIS evaluation framework with mutual feedback optimization， enabling multi-granularity satisfaction measurement from underlying performance metrics to top-level operational capabilities. Experimental case studies demonstrate that the proposed model effectively identifies weight inflation in cross-dependent indicators， achieves significant improvement in assessment objectivity， and reveals software capability deficiencies and operational constraints. This approach facilitates the evolution of software evaluation from "quality assurance" to "capability assurance".关键词：military software capability evaluation;hierarchical entropy-weighted TOPSIS;dynamic multi-task scenarios;demand satisfaction evaluation;equipment test and evaluation1|0|0更新时间：2026-08-03
- 摘要：With the deep development of astronomical observation technology toward multi-terminal collaborative intelligence， the limitations of traditional single-terminal applications in time-domain data integration， cross-platform compatibility， and user interaction have become increasingly prominent. Existing solutions rely on fixed architectures for device connection and data processing， where single communication protocols and limited computing power scheduling mechanisms struggle to meet the real-time collaboration requirements of high-throughput astronomical data.To address the above issues， propose and implement the Tianwenhong application system based on the HarmonyOS ecosystem. The system leverages HarmonyOS's distributed soft bus technology to network observation devices and user terminals， adopting the ArkUI declarative framework to develop adaptive UIs and address multi-screen inconsistency issues. It constructs a MongoDB database and utilizes the Node.js toolchain for end-cloud collaboration pipeline data processing， enabling energy-efficient storage of astronomical data and real-time visualization. By integrating the Analytic Hierarchy Process （AHP）， a comprehensive observation suitability model is established based on meteorological data， stellar attributes， and user interests to provide users with reasonable observation recommendations. The system also realizes one-code-multi-terminal deployment， HarmonyOS WebDatabase support， and compatibility with Android/iOS and other platforms. The system provides a full-process solution for astronomical observation covering device interconnection-data processing-intelligent interaction， promotes the innovative application of the HarmonyOS ecosystem in professional scientific computing scenarios， and establishes a reusable engineering paradigm for the deep integration of domestic information technology and astronomical informatization.关键词：HarmonyOS;ArkTS;astronomical observation;data visualization;Huawei cloud elastic cloud server1|0|0更新时间：2026-08-03
- 摘要：The data center carries the storage of critical business and massive data， and the equipment in the data center needs to operate stably 24/7. Therefore， it requires automated and integrated operation and maintenance. In order to facilitate operation and maintenance personnel to view equipment parameters and locate equipment faults， a power environment monitoring system based on DCOM architecture is proposed. Firstly， power， environment， and video surveillance data are collected through an embedded host and uploaded to the cloud； Secondly， through data preprocessing， monitoring and analysis， a three-dimensional visual operation and maintenance interface is generated to achieve automated monitoring and three-level alarms. In the practice of meteorological big data application center （Xi'an）， it was found that a total of 65535 alarms were generated in the past 6 months， including 750 severe alarms and 118 emergency alarms. The hit rate of severe and emergency alarms reached 100%， proving that the proposed system can efficiently coordinate monitoring， accurately locate faults， improve operation and maintenance efficiency， and ensure stable operation of the data center.关键词：operation and maintenance automation;DCOM architecture;power environment monitoring;computer room operation and maintenance1|0|0更新时间：2026-08-03
Software Design\, Development\, and Application
- 摘要：With the rapid advancement of healthcare informatization， medical images， which serve as essential carriers of patient privacy and critical diagnostic information， have become a focal point in encryption research and application. However， existing image encryption methods generally face challenges such as insufficient security， low efficiency， and limited functionality. To address these issues， this paper proposes a novel medical image encryption algorithm named FGC-RDH. Specifically， the algorithm first integrates a hash algorithm with a fractional-order 4D Chen hyperchaotic system to construct a unified pseudo-random number （PRN） generation mechanism， which drives dynamic fractional-order S-boxes and space-filling curve permutations to achieve efficient pixel scrambling. Second， in the diffusion phase， chaotic hash feature vectors are extracted from the plaintext as constraints to guide a Conditional Generative Adversarial Network （GAN） in generating a base diffusion key， which is then refined through adaptive chaotic processes for regionally enhanced diffusion. Finally， by incorporating plaintext-domain reversible watermarking technology， the algorithm enables reliable data traceability. Comparative experiments with algorithms such as RIE-RID， CCAES， and MCBS demonstrate that the proposed method significantly improves performance in terms of ciphertext entropy， NPCR， UACI， and key sensitivity. Notably， the NPCR reaches 99.616% and the ciphertext entropy achieves 7.999 1. These results prove that the algorithm possesses high security and robust anti-attack capabilities， providing strong technical support for the efficient and secure transmission of remote medical data.关键词：reversible watermarking;image encryption;machine learning;medical image1|0|0更新时间：2026-08-03
- 摘要：Cross-model generalization remains a critical bottleneck in synthetic image detection， primarily because images produced by diffusion models and GANs diverge significantly at both pixel-level statistics and high-level semantics. To bridge this gap， we introduce the Channel Variance-Guided Semantic-Pixel Feature Fusion Detection Method （CVGF）. Our framework adopts a dual-branch architecture for feature extraction. The pixel branch isolates low-level forgery fingerprints using a high-pass filtering strategy， while the semantic branch leverages a frozen CLIP model to capture inconsistencies in high-level content. The central novelty of our work is a channel variance-guided fusion module. Rather than simply concatenating features， this module models the statistical dispersion across heterogeneous feature channels and employs an extremely lightweight learnable base weight （
- 摘要：With the rapid development of IoT technology， medical data sensors are widely used in smart healthcare scenarios. While achieving medical convenience， the complexity of certificate issuance， verification， and storage procedures makes it vulnerable to attacks that can lead to patient privacy breaches. Certificate free signature is one of the key technologies to address security and privacy issues in smart healthcare scenarios. Aiming at the problems of complex sensor certificate distribution， vulnerability to centralized database attacks， and patient privacy leakage in existing certificate free signature schemes for smart healthcare， a blockchain based certificate free ring signature scheme BCLRS is proposed for smart healthcare. Firstly， deploy the traditional key generation center on the blockchain and combine elliptic curve public key cryptography with ring signatures to protect patient privacy； Secondly， achieve traceability of the identity of patients with special needs. Security analysis shows that the BCLRS scheme meets the requirements of correctness and anonymity， and can resist malicious attacks such as man in the middle attacks， distributed denial of service attacks， and key generation center compromise attacks compared to unlicensed circular signature schemes. Performance analysis and simulation experiments show that when the number of ring members exceeds 30， the computational cost of the BCLRS scheme is reduced by 80%， making it suitable for resource constrained smart healthcare scenarios.关键词：smart healthcare;certificateless ring signature;blockchain;random oracle model;privacy-preserving1|0|0更新时间：2026-08-03
Cyberspace Security
- 摘要：To address the challenges posed by financial domain corpora—characterized by diverse entity types， complex relationship structures， dense domain-specific terminology， and subtle semantic distinctions—this paper proposes a relation classification method that integrates N-way K-shot few-shot learning with a multi-stage prompting mechanism for large language models（LLMs）. Based on financial text corpora， the method constructs a support set and a query set， leveraging a pretrained language encoder to select semantically similar examples. These examples are embedded into structured prompts to guide the LLM in relation recognition. During inference， an adaptive verification mechanism is introduced to perform self-assessment and correction of the model's initial predictions， thereby enhancing the stability and accuracy of the results. Experimental results demonstrate that the proposed method significantly outperforms traditional few-shot approaches in terms of precision and recall across several mainstream LLMs. Moreover， the method maintains robust performance under zero-shot settings， highlighting its practical value and strong generalization capability in financial relation extraction tasks.关键词：information extraction;relation extraction;few-shot learning;zero-shot learning;large language model1|0|0更新时间：2026-08-03
- 摘要：Formal Concept Analysis （FCA） explores interpretable rules in a structured manner and has broad application prospects. This paper proposes a novel rule classification model based on linguistic concept lattice（LCLC）， which extracts semantic classification rules based on linguistic concept lattice. First， input data is mapped to linguistic terms via fuzzy partitioning to build a linguistic concept formal context， from which the linguistic concept lattice is derived. Next， a rule extraction algorithm based on linguistic concept lattice is proposed， which determines the rule base for classification by considering the concept area and rule weight. Then， the rule reasoning mechanism of LCLC is introduced， where the antecedent membership degree of the rule is calculated using a weighted average of the membership degrees after fuzzification of the input， while considering the rule weights to determine the classification label. Ultimately， the effectiveness of LCLC was validated through comparative evaluations with machine learning classification and other rule-based classification algorithms across eight public datasets，that LCLC achieved a classification accuracy of 100% and had a more concise set of rules.关键词：formal concept analysis;classification rule;linguistic concept lattice;knowledge discovery;rule extraction1|0|0更新时间：2026-08-03
- 摘要：Aiming at the problems such as the complex strokes and high structural similarity of Tangut characters， as well as the existence of missing characters and fading in Tangut ancient books， resulting in unsatisfactory detection accuracy， missed detection and incorrect detection of characters， a Tangut character detection method based on SCA-DBNet is proposed by comprehensively analyzing the existing mainstream character detection methods. This method first introduces the Spatial and Channel Reconstruction Convolution module （ScConv） to replace the original convolution module （Conv） and combines it with the MobileNetV3 feature extraction network to reduce computational redundancy and improve the feature expression ability of the model. Meanwhile， in the feature fusion stage， the Adaptive Scale Fusion module （ASF） is introduced to solve the problem of inconsistent scales of features in different layers during the feature fusion process， and the Hybrid Attention module （CFAM） is designed to replace the original Spatial attention module （SAM） to enhance the characterization of multi-scale varying features， thereby improving the accuracy of text detection in Tangut ancient books. Compared with the basic model on the self-built Tangut text dataset， the accuracy rate and recall rate have increased by 3.2% and 2.9% respectively. Compared with other text detection models， there have been certain improvements in both accuracy rate and recall rate， and it can accurately detect Tangut text.关键词：text detection;spatial and channel reconfiguration convolution;adaptive scale fusion;attention mechanism;Tangut ancient books1|0|0更新时间：2026-08-03
- 摘要：With the rapid advancement of artificial intelligence technology， the complexity of data relationships has expanded from traditional two-dimensional structures to multi-dimensional ones， posing significant challenges for conventional formal concept analysis in handling multi-dimensional data. To address this issue， this paper proposes a fuzzy linguistic triadic concept analysis framework， aiming to meet the analytical demands of complex multi-dimensional data in uncertain environments. First， by integrating linguistic term sets with triadic concept analysis， this paper constructs a fuzzy linguistic triadic context. Based on the conditional concept-forming operator， this paper introduces an algorithm for generating fuzzy linguistic triadic concepts. Second， this paper constructs a fuzzy linguistic triadic decision context and introduces a rule extraction algorithm with confidence measures， enabling the discovery of more latent knowledge from multi-dimensional data. Finally， by applying the RFM model to analyze Taobao user shopping behavior data， this paper validates the effectiveness and rationality of the proposed method.关键词：fuzzy linguistic triadic context;linguistic term set;knowledge discovery;rule extraction;RFM model1|0|0更新时间：2026-08-03
- 摘要：With the rapid development of the internet， people are in an era of information overload. Understanding users' dynamic preferences through conversational and providing personalized recommendations is of great significance. To address the issues of specific knowledge scarcity and hallucination in conversational recommendation systems using large language models （LLM）， this study proposes an innovative method based on Knowledge Graph Retrieval-Augmented Generation. The core of this method lies in transforming knowledge graphs into natural language documents and dynamically retrieving relevant knowledge based on user-system dialogues， which is then integrated into predefined templates to generate accurate recommendations. This approach effectively aligns the semantic spaces of knowledge graphs and LLM and enriches LLM with extensive domain knowledge， thereby significantly mitigating hallucination issues. Furthermore， by seamlessly incorporating recommendation results into LLM， the method achieves coordinated responses between the dialogue module and the recommendation module. Experimental results demonstrate that the proposed method outperforms existing approaches in terms of recommendation accuracy and dialogue response quality on two public datasets. Ablation studies further validate the effectiveness of each component in the method.关键词：conversational recommender systems;knowledge graph;large language models;retrieval-augmented generation;natural language processing1|0|0更新时间：2026-08-03
Big Data and Information Processing
-
ConoHN： A Hierarchical Knowledge-Enhanced Framework for Fine-Grained Conodont Image Classification AI导读
摘要：Conodont identification is a critical task in paleontology， yet current image classification methods remain limited by insufficient feature learning and difficulties in capturing discriminative multi-region variations across species. To address these challenges， we propose ConoHN， a hierarchical multi-label classification framework that incorporates taxonomic knowledge. Our key innovation decomposes the complex task of locating sub-category discriminative regions into simpler hierarchical subtasks through attention-based feature discretization across taxonomic levels （family-genus-species）， and high-level feature erasure to force learning of low-level discriminative patterns. Experiments demonstrate that ConoHN significantly outperforms state-of-the-art methods in multi-level classification accuracy， while providing interpretable region-focused attention maps.关键词：conodont;fossil image classification;hierarchical multi-label image classification;fine-grained image classification1|0|0更新时间：2026-08-03 - 摘要：Railway tracks are the basic infrastructure that ensures the operation of trains. As the operating mileage continues to increase， irreversible damage may occur on the surface of the tracks， which can increase the risk of train operation in severe cases. For this purpose， an improved YOLOv11n railway track surface damage detection algorithm was designed. Firstly， using the Imageio library to increase snow and fog rendering data can improve the model's generalization ability and training efficiency； Secondly， integrating lightweight dynamic convolution into the backbone network significantly enhances the feature extraction capability of rail damage； Once again， establish the SCSA mechanism to achieve the fusion of backbone features， fully utilize the multi-level semantic information in the model， and optimize the performance of model detection tasks； Finally， utilizing the fine-grained deep structure in the ADown downsampling module improves the capture efficiency of spatial features and reduces additional computational burden. Experiments on the rail damage dataset show that the mAP50 of the DSA-YOLO algorithm reaches 98.1%， and the parameter count is reduced to 2.29 × 10 ⁶， which is superior to the comparative algorithms. The generalization experiment on the public steel dataset showed that the DSA-YOLO algorithm improved the mAP50 index by 5% compared to the comparison model， proving that the proposed model has strong generalization on different datasets.关键词：rail surface damage detection;YOLOv11n;data augmentation;spatial-channel coordination;dynamic convolution1|0|0更新时间：2026-08-03
- 摘要：Meteorological image resolution is crucial for studying climate phenomena. Although meteorological over-sampling methods based on deep learning have made rapid development， and the ability to learn low-resolution image features has been continuously strengthened. However， the up-sampling of current methods is mainly based on inverse convolution， interpolation and sub-pixel convolution； all three methods cause certain artifacts， and features are lost during up-sampling resulting in poor reconstruction results. In order to better utilize the low-resolution features in upsampling and thus improve the quality of upsampling， this paper proposes a meteorological supersampling algorithm based on residual attention and dynamic upsampling - RANSR， including a low-resolution feature extraction module， a dynamic upsampling module and a high-resolution feature fusion module. Among them， the low-resolution feature extraction module strengthens feature learning based on deep residual attention network， the dynamic up-sampling module generates more accurate up-sampled features and enhances the model generalization ability through the content-aware restructuring mechanism， and the high-resolution feature fusion module fuses the auxiliary inputs including elevation and climate state to enhance the final reconstruction effect. In the rainfall 4-fold， rainfall 10-fold， 2m temperature 4-fold， and 2m temperature 10-fold supersampling experiments， comparing the five meteorological supersampling methods， the MSE is reduced by 32.0%， 21.4%， 27.3%， and 55.9%， respectively， and PSNR is improved by 4.2%， 3.3%， 4.2%， and 12.1%， respectively， compared with the method with the second best performance.关键词：super-resolution;deep learning;up-sampling;attentional mechanisms;precipitation1|0|0更新时间：2026-08-03
- 摘要：This paper proposed an uncertainty modeling method based on the YOLOv11 object detection algorithm to address the challenges posed by rainy， snowy， and foggy weather to the accuracy and reliability of autonomous driving visual perception systems. The method used a Laplace distribution to model the uncertainty of predicted bounding boxes and employed information entropy to quantify the uncertainty of classification tasks. These quantified uncertainties were incorporated into redesigned bounding box loss functions and classification loss functions， enabling the model to adaptively adjust its focus on bounding box prediction accuracy and classification confidence. Experimental validation was conducted on the CADC， BDD100K， and a self-built dataset， demonstrating that the proposed method achieves improvements of 9.3%， 5.2%， and 8.6% in mAP@0.5 compared to baseline models. The experimental results show that the method significantly improves the robustness and accuracy of algorithms under adverse weather conditions， reduces the probability of missed or false detections， and enhances detection precision and prediction reliability.关键词：autonomous driving;object detection;YOLOv11;uncertainty modeling;Laplace distribution1|0|0更新时间：2026-08-03
- 摘要：The early diagnosis of pulmonary nodules is essential to improve the survival rate of patients. However， traditional detection methods are prone to lose small target features， difficult to resist complex background interference， and lightweight and performance imbalance. Therefore， a lung nodule detection algorithm based on improved YOLOv9 is proposed. Firstly， the loss function Shape IoU is introduced to enhance the model's attention to complex backgrounds and small targets by optimizing the traditional calculation method of intersection and merger ratio； Secondly， the space channel collaborative attention module is designed， and multi-scale spatial feature fusion and progressive channel compression strategy are used to enhance the feature expression ability of the algorithm； Finally， the lightweight dynamic convolution module is designed. Through the dynamic selection strategy of expert convolution core library， the flexibility of feature extraction is improved while maintaining the lightweight of the model. The experimental results show that the average accuracy （AP） of the improved algorithm is 82.7%， which is 2.4% higher than the original YOLOv9 model. The accuracy and recall rates are 2.8% and 0.6% higher respectively， and the number of parameters is reduced by 5 × 106. This algorithm realizes the balance between detection accuracy and lightweight， and provides an efficient solution for the detection of pulmonary nodules.关键词：lung nodule detection;YOLOv9;lightweight dynamic convolution;Shape-IoU;spatial-channel collaborative attention1|0|0更新时间：2026-08-03
- 摘要：Learning-based multi-view stereo （MVS） has achieved notable progress， yet remains challenged by textureless regions， severe occlusions， and complex geometric structures. To tackle these limitations， this paper introduces SM-MVS， a novel framework that synergistically integrates curvature-guided feature enhancement with epipolar geometry-based iterative refinement. The proposed approach consists of two key components： a multi-scale curvature-aware network that adaptively modulates convolutional kernels using geometric priors to reinforce edge features while suppressing noise， and an epipolar optimization module that progressively refines depth estimation through dynamic sampling and cross-view attention. Comprehensive evaluations on the DTU dataset demonstrate that SM-MVS achieves a 22.1% improvement in completeness and an 11.5% increase in overall score compared to MVSNet， while maintaining competitive computational efficiency. These results validate the framework's robustness and practical utility for high-fidelity 3D reconstruction in complex scenarios.关键词：deep learning;epipolar geometry;multi-view stereo;3D reconstruction1|0|0更新时间：2026-08-03
Graphic and Image Processing
- 摘要：Image denoising is a fundamental step in digital image processing and plays a crucial role in improving image quality and enabling subsequent processing.Constrained by factors including the precision limitations of imaging devices and transmission environment interference， various mixed noise artifacts generated during the imaging process lead to image blurring， compromised detail resolution， and chromatic distortion. These degradations persistently affect both daily life and professional activities， while the growing demand for acquiring higher-precision and clearer images makes image denoising particularly crucial.Based on denoising algorithm typology， these methods are categorized into filtering-based approaches， low-rank optimization， sparse optimization， and deep learning-based frameworks. Their performance is systematically analyzed in terms of processing speed， noise suppression capability， and edge preservation characteristics， with comparative evaluations conducted through quantitative metrics such as PSNR.A comprehensive evaluation of denoising performance across methodologies is conducted for diverse image types， systematically benchmarking their strengths and limitations. The analysis further delineates applicable scenarios and development trajectories for each approach， with particular emphasis on emerging frontiers including real-time denoising for dynamic imagery， parameter self-adaptation mechanisms， and data-efficient paradigms that circumvent reliance on large-scale training datasets.关键词：image denoising;digital image processing;edge preservation;deep learning;real-time denoising1|0|0更新时间：2026-08-03
- 摘要：As a swarm intelligence algorithm derived from the social hierarchy and cooperative hunting behavior of gray wolves， gray wolf optimization algorithm has received extensive attention and in-depth research in recent years due to its simple structure， few parameters， easy implementation and other advantages. The analysis is carried out around the grey wolf optimization algorithm. First， we introduce the biological background. Then， analyze the mathematical modeling process of the algorithm. Finally， summary the main improvement methods and applications of the algorithm systematically. These provide a reference frame for the advanced theoretical and application of the grey wolf optimizer.关键词：swarm intelligence;grey wolf optimizer;machine learning;engineering optimization1|0|0更新时间：2026-08-03
- 摘要：Knowledge graph is an important research direction in the field of artificial intelligence， and one of its core tasks is knowledge graph completion. This task encompasses triple classification， link prediction， and relation prediction， aiming to fill in missing head entities， tail entities， or relations. With the rapid development of large language models in the field of natural language processing， research on the integration of knowledge graph completion with large language models has received increasing attention， and various models that can effectively handle graph completion tasks have emerged. Firstly， the basic concept of knowledge graph completion is elaborated； secondly， knowledge graph completion models integrated with large language models are discussed； thirdly， common datasets and evaluation methods for knowledge graph completion models based on large language models are introduced； finally， the challenges of knowledge graph completion models based on large language models are summarized， and future development directions in this field are prospected. The research results can provide a comprehensive reference for the current status of this field for relevant researchers， help subsequent research to further break through the limitations of existing methods， and promote the development and application of knowledge graph completion technology.关键词：large language model;knowledge graph completion;natural language processing1|0|0更新时间：2026-08-03
Research Review
- Technical support is provided by Beijing Founder electronics co., LTD 鄂ICP备18021494号-4
鄂公网安备42010602004949号 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0