ABSTRACT To address inaccurate fruit recognition and difficult coordination between grasping and harvesting region localization for cucumber harvesting robots in complex greenhouses, a vision system from instance segmentation to dual‐region collaborative localization is proposed. First, the LPG‐YOLO11n‐seg model is built upon YOLO11n‐seg; a lightweight instance segmentation head LIS‐SB improves illumination robustness, a parallel patch‐aware attention module PPA enhances feature extraction, and GFPN with CSPStage structures optimizes multi‐scale feature fusion to achieve high‐precision instance segmentation of cucumber fruits. Then, a dual‐region collaborative localization method is proposed: geometric parameters are extracted from segmentation masks; the optimal grasping point is determined by comparing distances from the stem‐tip midpoint and the maximum inscribed circle center to the centroid; and a gripper‐matched grasping region is generated. Through scale adjustment and safety offset strategies, the grasping region is converted into a tool harvesting region that avoids the fruit body, realizing geometric collaboration between grasping and harvesting regions. LPG‐YOLO11n‐seg improves mAP50 by 6.1%, 9.7%, 3.1%, and 6.0% over YOLOv5n‐seg, YOLOv9t‐seg, YOLO11n‐seg, and YOLO12n‐seg, respectively. For instance segmentation, it achieves 92.5% precision, 88.8% recall, 95.1% mAP50, and 76.0% mAP50‐95, all significantly higher than YOLO11n‐seg. The average center‐point deviations of grasping and harvesting regions are 2.19% and 2.20%, with high‐precision localization rates exceeding 95%. Embedded deployment on the low‐power platform optimized by TensorRT demonstrates excellent efficiency, a network inference speed of 44.66 FPS and a stable continuous end‐to‐end on‐board processing latency of 434.0 ms. The system provides a reliable visual localization solution for cucumber harvesting robots in complex greenhouse environments.
ABSTRACT Decentralized scholarly publishing offers improved transparency and long‐term preservation of scientific knowledge, yet most existing approaches rely on off‐chain storage for article content, introducing external dependencies and weakening availability guarantees. This paper presents a practical approach for fully on‐chain storage of canonical LaTeX article content. To achieve cost‐efficient deployment, an inscription pipeline is designed that integrates compression, hybrid encryption, and chunk‐based storage, where textual content is compressed to reduce on‐chain data footprint, encrypted using symmetric encryption with asymmetric key protection, and partitioned to balance cost efficiency and execution constraints. External assets are stored in an Ethereum‐compatible decentralized storage layer using a path‐preserving mechanism, enabling deterministic reconstruction without additional metadata. Experimental evaluation on real‐world articles, including deployments on a public Ethereum test network, demonstrates that the proposed approach achieves reliable on‐chain storage with manageable cost and efficient reconstruction, providing a feasible solution for fully on‐chain scholarly publishing. The implementation is publicly available at https://github.com/thales‐must/dapp‐ri.git .
ABSTRACT In recent years, quantum emulation has attracted increasing attention as a practical approach for developing and evaluating quantum algorithms. However, existing solutions often face challenges in achieving broad applicability, high performance, and efficient utilization of hardware resources and memory. To address these limitations, this paper presents the high‐performance quantum emulation accelerator (HPQEA), a hardware architecture based on the state‐vector emulation approach. The proposed design incorporates four key techniques: a strict data storage strategy, a superior flexible processing element network, an optimized and parallelized CX swapper, and a high‐bandwidth memory‐based bulk data transfer engine. Implementation and evaluation on the Alveo U280 FPGA platform demonstrate that HPQEA can emulate quantum circuits efficiently while maintaining high fidelity and low mean square error. Experimental results show that for QFT benchmarks, HPQEA achieves up to 21,093 faster execution than the NVIDIA A100 GPU for circuits with fewer than 20 qubits. Moreover, the comparison with FPGA‐based related works shows promising results with significant speedup in terms of normalize gate speed from 2.77 to 1448.37. In addition, the architecture demonstrates strong energy efficiency, reducing power consumption by up to 15.58 compared to the Intel i9‐10940X CPU and 80.83 compared to the GTX 1080 GPU. These results highlight the scalability, efficiency, and practicality of HPQEA as a hardware platform for accelerating quantum algorithm emulation.
ABSTRACT Hierarchical Edge–Fog–Cloud infrastructures increasingly support distributed sensing, analytics, and collaborative computation over sensitive data. Their multi‐tier structure creates systems challenges for secure orchestration: resources are heterogeneous, communication costs vary sharply across layers, and trust assumptions are often bounded by administrative domains. This paper presents H‐SPDZ‐Cloud, a formal prototype for hierarchical authenticated secure multi‐party computation over Edge–Fog–Cloud architectures. The design organizes computation into cryptographically isolated levels, each using an independent message authentication code key, and introduces a verified inter‐level resharing mechanism that keeps secrets in shared form during normal‐case upward transitions. The prototype also includes localized fault handling under a per‐level honest‐majority assumption suited to bounded hierarchical deployments. The H‐SPDZ‐Cloud prototype is evaluated in a controlled single‐host reproducibility testbed configured with the 127‐bit Mersenne prime field , represented using 16‐byte field‐element encodings. Prototype measurements show communication‐locality behavior relative to an internal flat SPDZ‐style Python baseline, with a 31.4% reduction in measured online communication and a 42.4% reduction in total communication for the tested workload. The prototype also evaluates secure fixed‐point inference on the public Wisconsin Diagnostic Breast Cancer dataset, with secret‐shared test features and 100% agreement between the secure and quantized plaintext predictions on the secured subset. The implemented 127‐bit‐prime‐field pipeline uses fixed‐base precomputation for the BN254/Pedersen transition evidence, the corrected masked batch MAC check, and a prototype Sigma/Schnorr dispute‐proof fallback, yielding a measured local‐pipeline latency of 696.2 ms and an fault‐localization latency of 3066.8 ms as reproducible prototype timings. TCP validation with independent party servers further exercises socket serialization and process‐isolated Edge/Fog transition workers; a subsequent authenticated multi‐region three‐VPS validation executes the external‐endpoint path from an independent Cloud controller to Edge and Fog VPS workers over TLS 1.3, reporting 229.2 ms mean TCP/TLS latency, 93.5 ms transition‐fault detection with 30/30 detections, and 264.3 ms/sample secure‐inference latency. A separate local MP‐SPDZ MASCOT reference execution confirms that the three reference arithmetic workloads also run in a standard MPC framework, while H‐SPDZ‐Cloud measurements use the Python prototype pipeline. The manuscript implements and measures an optimistic verification and pessimistic privacy‐preserving dispute path, with plaintext audit opening reserved for a separate administrative debug mode.
ABSTRACT Cryptocurrencies and blockchain have enabled trustless, low‐cost cross‐border transactions and promoted the rapid development of decentralized e‐commerce. However, existing decentralized e‐commerce systems face three critical limitations: insufficient privacy‐preserving mechanisms, lack of Arbitratory accountability, and unfair dispute resolution without victim compensation. These issues severely hinder real‐world adoption. To address these challenges, this paper proposes a novel regulated decentralized e‐commerce transaction framework based on dual‐escrow and multisignature smart contracts, integrating game‐theoretic deterrence, cryptographic privacy protection, and conditional identity disclosure. Our main contributions are threefold: (1) Methodological Innovation: We propose a novel compensating MAD game‐theoretic mechanism that mathematically guarantees that violators incur losses exceeding their illegal gains while ensuring honest parties receive full compensation—a property not achieved by prior MAD models. (2) A decoupled contract architecture using 2‐of‐2 multisignature for fund control and 2‐of‐3 multisignature for arbitration, with native design support for BLS aggregate signature extension to reduce gas costs and improve scalability in multiarbitrator scenarios; (3) A privacy‐preserving real‐name commitment exchange protocol based on cryptographically binding real‐name commitments to transactions, enabling Arbitrator‐accessible identity disclosure only upon verified violation; the architecture further reserves modular interfaces for zk‐SNARK integration to conceal transaction amounts and addresses without sacrificing verifiability. We implement the system as smart contracts on the Ethereum testnet and conduct comprehensive security analyses and performance evaluations. The results demonstrate that our scheme achieves strong decentralization, low fraud risk, compliant accountability, lower gas consumption, and reduced on‐chain storage overhead compared with state‐of‐the‐art solutions. This work provides a practical and secure paradigm for real‐world deployable decentralized e‐commerce systems. Meanwhile, a concurrency scheduling strategy is proposed to resolve state read‐write conflicts and lock contention of smart contracts under massive parallel transaction requests to optimize system throughput and latency in high‐concurrency scenarios.
ABSTRACT The Maximum Satisfiability (MaxSAT) problem is a fundamental combinatorial optimization task, and state‐of‐the‐art solvers rely on decades of expert‐engineered heuristics and low‐level micro‐optimizations to achieve competitive performance. While recent Large Language Model (LLM)‐driven approaches (e.g., FunSearch, SATLUTION) have shown promise in algorithmic design and global solver evolution, they either focus on high‐level algorithmic sketches or large‐scale code restructuring, leaving the potential of purely local, correctness‐validated code optimization for industrial MaxSAT solvers largely underexplored in prior MaxSAT‐focused studies. This paper presents LoMax, a plug‐and‐play framework that leverages an LLM to perform local code optimization for MaxSAT solvers. Without any human‐written heuristics or instance‐specific domain‐specific templates (The prompts contain general MaxSAT domain knowledge but no instance‐specific or benchmark‐specific optimization templates), LoMax iteratively (i) ranks source files, (ii) scores individual functions, (iii) optimizes the highest‐scoring function body, (iv) performs a fast sanity evaluation on a small instance set, and (v) validates it on full exact‐unweighted and exact‐weighted benchmarks. Experiments on two state‐of‐the‐art solvers (UWrMaxSat‐SCIP‐MaxPre and OpenWBO) show consistent improvements: UWrMaxSat‐SCIP‐MaxPre solves additional instances (average runtime 13.11 s 14.27 s) on 553 unweighted instances and additional solved instances (average runtime 11.67 s 10.76 s) on 571 weighted instances, both under a uniform 100‐second timeout; OpenWBO solves additional instances ( average time) on 553 unweighted instances, providing empirical evidence that LLM‐driven local code rewriting can improve the performance of industrial MaxSAT solvers in the evaluated settings.
ABSTRACT Deep learning‐based vision models have severe security vulnerabilities against physical adversarial patch attacks. Previous studies propose explainability‐based defense mechanisms (e.g., PatchOut and SALIUITL). These methods aim to recover the system by masking the regions that affect model decisions the most. However, these approaches generally assume that the adversarial effect is located in a single, continuous region (single mask selection). In this study, we empirically demonstrate the “Semantic Fragmentation” and saliency bleeding problems caused by multiple scattered patches in explainability maps. To overcome this issue, we propose the “Top‐K Semantic Mask Fusion with Dynamic Budget (TK‐SMF)” system. The proposed framework initially extracts explainability maps using LayerGradCAM, subsequently constructs a pool of candidate semantic segments via the Fast Segment Anything Model (FastSAM), and ultimately fuses the Top‐K highest‐risk masks within a strictly fixed area budget. This budget is anchored to a predefined worst‐case threat assumption, eliminating the need to know the attacker's actual patch count during inference. We then restore the selected masked areas using OpenCV Inpainting to achieve model recovery. We evaluated the proposed framework on 38 distinct and challenging physical‐world scenarios, encompassing diverse tactical vehicle categories under single‐ and multipatch attacks. Experimental results prove that TK‐SMF breaks the targeted attack illusion with a 100% success rate. The results show that dynamic and flexible defense architectures exceed the limits of the single‐mask assumption. They significantly improve model accuracy and reliability against complex physical world threats.
ABSTRACT With the advancement of deep learning, diffusion models have gained prominence as a powerful method for high‐fidelity image generation by progressively removing noise to restore structural and textural details. In remote sensing image super‐resolution (RSISR), which typically utilizes paired high‐resolution (HR) and low‐resolution (LR) images for supervised training, applying diffusion models directly faces three persistent challenges. First, common preprocessing techniques such as bicubic interpolation lose substantial prior information, resulting in a lack of high‐frequency details and weakened spatial structural features. Second, existing denoising networks often exhibit insufficient global modeling capability, performing poorly in capturing long‐range spatial relationships and evaluating channel‐wise feature importance. Third, traditional noise prediction (‐prediction) struggles to restore fine details under high‐noise conditions, frequently failing to preserve the complex structures of ground objects in remote sensing scenes. To address these issues, this paper proposes DAV‐Diff, a fast‐sampling conditional diffusion model for RSISR. DAV‐Diff coordinates three task‐oriented components to improve conditional prior representation, long‐range dependency modeling, and high‐noise detail recovery in RSISR. The first is a multi‐dilated residual attention block (MD‐RAB), which enhances the extraction of prior information from LR images and compensates for detail loss caused by interpolation. The second component is a dual‐attention mechanism embedded within the denoising network to strengthen long‐range spatial dependency modeling and adaptive channel feature weighting. The third is a residual‐domain ‐prediction strategy that replaces traditional noise prediction, mitigating detail degradation in high‐noise environments while preserving structural features of complex ground objects. Experimental results show the effectiveness of DAV‐Diff against state‐of‐the‐art methods; ablation studies, computational complexity analysis, and sampling‐step sensitivity experiments further quantify the contribution and efficiency of each component.
ABSTRACT Predicting the runtime and cost of Function‐as‐a‐Service (FaaS) applications remains challenging in multi‐cloud environments due to variations in code complexity, workload characteristics, and provider‐specific behaviors. This paper presents an extended version of the Orama Framework that advances runtime prediction toward a multilingual and cost‐aware approach. The framework incorporates a multilingual Halstead metric extractor for language‐agnostic static analysis and enhances the predictor to estimate execution costs by combining runtime forecasts with cloud pricing models across AWS Lambda, Google Cloud Functions, Azure Functions, and Alibaba Function Compute. To assess robustness and generalization, the data set is expanded with a scientific workload based on genetic sequence alignment, introducing input‐sensitive execution patterns. The extended data set integrates static code metrics, workload scale, infrastructure metadata, and empirical multi‐cloud execution traces. Neural network models (Dense, LSTM, and BLSTM) are retrained and evaluated using standard regression metrics and cost estimation. Results indicate that the enhanced BLSTM model maintains high predictive precision across heterogeneous workloads and providers, while enabling cross‐cloud cost estimation directly from source code. The extended framework provides a unified approach for performance and cost‐aware prediction in serverless computing environments.
ABSTRACT Robust representation‐based image classification must cope with complex mixed degradations, including global structural variations (e.g., severe non‐linear illumination) and localized contiguous corruptions (e.g., block occlusions or disguises). While existing matrix‐based low‐rank regression methods effectively model structural residuals without breaking the intrinsic two‐dimensional spatial correlations of pixels, they consistently neglect the underlying geometric relationships among training samples. This oversight leads to dense and weakly discriminative representation coefficients, particularly in under‐determined small sample size (SSS) scenarios. To address this limitation, we propose the Manifold Regularized Nuclear Norm Matrix Group Sparse Classifier (MNMGSC), a unified convex matrix optimization framework. MNMGSC jointly enforces a low‐rank residual via a nuclear‐norm penalty, explicitly models contiguous occlusion via a matrix structural penalty, promotes class‐wise discriminative coefficients via a group‐sparse ‐norm, and ensures geometric consistency via a Graph Laplacian regularizer. This mathematical design explicitly yields a formal ‐spectrum that bridges sparse and collaborative representation paradigms. To efficiently solve the proposed non‐smooth model, we develop a Nesterov‐accelerated Alternating Direction Method of Multipliers (ADMM) and provide explicit theoretical guarantees, including a Group Sparse Low‐Rank Restricted Isometry Property (GS‐LRIP) recovery error bound and convergence proof. Evaluated extensively on the ORL, Extended Yale B, AR Face, and MNIST databases against both the unregularized baseline NMGSC and recent state‐of‐the‐art matrix/tensor representation methods, MNMGSC demonstrates statistically significant performance improvements (), particularly under severe non‐linear degradations and highly limited training samples.
ABSTRACT Surface electromyography (sEMG) signals are vital bioelectrical indicators for decoding human motion intentions in rehabilitation robotics. However, achieving high‐precision synchronous modeling of hand motion pattern recognition and joint angle prediction remains challenging. This study proposes a shared‐backbone sEMG‐based modeling framework built upon the MobileViT backbone, which integrates the local feature extraction capability of convolutional structures with the global dependency modeling of Transformers. An improved energy kernel method for the two‐dimensional (2D) characterization of temporal data is developed to generate interpretable 2D characterization of muscle activity through matrix‐based quantification of amplitude–velocity distributions. In addition, a feature discriminant analysis approach is introduced to visualize and evaluate the separability and intra‐class compactness of deep features across model layers. Using the NinaPro DB7 dataset, two task‐specific models were constructed: MobileViT‐PR for motion pattern recognition and MobileViT‐MP for finger joint angle prediction. Experimental results demonstrate that the proposed framework achieved recognition accuracies of 99.8% and 99.7% for intact and amputee subjects, respectively. In regression tasks, the MobileViT‐MP model yielded an average coefficient of determination R‐Square of 0.925 and a root mean square error of 6.444. Furthermore, the effects of multiple factors—including window length, stride, exercise type, subject age, and condition—were systematically analyzed to provide comprehensive references for future sEMG modeling and rehabilitation studies. Overall, the proposed framework enables dual‐task modeling with high interpretability, discriminative capability, and computational efficiency, offering an effective solution for human–robot interaction and rehabilitation control.
ABSTRACT This paper proposes a Hierarchical and Stagnation‐Aware Particle Swarm Optimization with Elite‐Guided Crossover (HECO‐PSO). HECO‐PSO dynamically partitions the swarm into multiple parallel layers, enabling independent exploration and periodic information exchange to mitigate premature convergence. Within each layer, elite‐guided crossover integrates a particle's best position with that of the layer's elite particle, generating high‐quality guiding vectors to enhance solution quality and maintain diversity. Building on this hierarchical framework, a stagnation‐aware strategy adaptively intensifies guidance: prolonged stagnation activates crossover with the global elite, while severe stagnation directly assigns the global best guiding vector, thereby strengthening the ability to escape local optima. Experimental results on the CEC 2017 benchmark suite demonstrate HECO‐PSO's competitive convergence speed and solution accuracy. Its effectiveness is further validated through a real‐world UNSW‐NB15 feature selection task, confirming adaptability and competitiveness in practical applications.
ABSTRACT One of the main challenges with influence maximization on complex networks is identifying a small yet effective set of influential spreaders. Current centrality‐ and voting‐based techniques tend to choose nodes within the same dense parts, resulting in overlapping influence and low spreads. Additionally, they are based on predetermined heuristics and cannot adjust to various network structures or historic diffusion patterns. To solve these problems, we introduce Improved CVoteRank, a learning and community‐aware instructional structure of voting that employs the information given by earlier states of the network to inform seed selection. The approach is a concatenation of several structural characteristics that symbolize both global and local spreading capacity, such as size‐sensitive Kullback‐Leibler divergence, participation coefficient, and intra‐community betweenness centrality. The ridge regression model is trained on the previous results of the diffusion and is automatically trained to find the most appropriate significance of the features. The resultant learned weights are used in a refined iterative voting process using concave neighbor influence and two‐hop suppression to minimize redundancy and enhance seed diversity. Experimental results on seven real networks in social, collaboration, infrastructure, and peer‐to‐peer systems indicate that Improved CVoteRank can always achieve faster information spreading and larger ultimate outbreak sizes of the Susceptible‐Infected‐Recovered (SIR) model. The proposed approach also shows a high level of robustness when the seed budgets and rates of infection vary. Structural analysis also indicates that the chosen spreaders are more uniform, and there is more space between seed nodes. These results validate that the adaptive voting approach in the form of learning has a significantly higher performance compared to the heuristic approaches that are fixed and thus offers an effective influence maximization tool in complex networks.
ABSTRACT Accurate delineation of breast lesions in ultrasound images can support reproducible lesion measurement and quantitative analysis after lesion presence has been established. Although convolutional neural networks (CNNs) have achieved promising performance in medical image segmentation, their limited receptive fields hinder long‐range contextual modeling. Transformer‐based methods address this limitation through self‐attention. However, conventional dense attention often introduces redundant token interactions and remains susceptible to ambiguous background responses in BUS images. In addition, generic feed‐forward transformations do not explicitly promote spatial consistency within lesion regions. To address these challenges, we propose BGLC‐Net, a boundary‐guided and lesion‐consistent CNN‐Transformer network for BUS lesion segmentation. Its core module, the Lesion‐Guided Transformer Block (LGTB), is embedded in each encoding stage and comprises Boundary‐Guided Token Selection (BGTS) attention and a Lesion‐Region Consistency Feed‐Forward Network (LRC‐FFN). BGTS incorporates boundary‐related guidance to emphasize contour‐informative responses while suppressing background interactions. LRC‐FFN strengthens lesion representation through progressive multi‐scale contextual modeling, improving lesion‐region consistency across multiple contextual scales. Experiments on the BUSI dataset show that BGLC‐Net achieves a Dice score of 81.02%, a Jaccard index of 73.12%, and an HD95 of 8.84. Overall, these results are superior or competitive compared with those of representative BUS segmentation methods. External validation on the independent STU dataset further suggests preliminary cross‐dataset generalization under dataset‐level appearance shifts, indicating that BGLC‐Net may provide a useful technical basis for reproducible BUS lesion segmentation.
ABSTRACT Accurate multivariate ocean wave forecasting requires representing rapidly evolving local dynamics, longer‐term background variation, and dependencies among sea‐state variables. This study proposes a Gated Cross‐Scale Residual Fusion Network (GCRF‐Net) for multivariate one‐step ocean wave forecasting. The 120‐step short‐window main branch uses long short‐term memory (LSTM), convolutional temporal, and attention encoders to form the base prediction. A 960‐step auxiliary LSTM branch generates a candidate residual correction. A variable‐wise gate conditioned on the main‐scale and long‐scale representations then regulates this correction. Under a station‐specific evaluation protocol, experiments on three independently processed Queensland wave‐buoy stations showed that GCRF‐Net achieved the lowest mean one‐step mean squared error (MSE) among the evaluated models at each station. Direct multi‐horizon results were horizon‐ and station‐dependent. This finding suggests that the model is most directly suited to the intended one‐step (0.5 h) nowcasting task. Ablation studies supported the contributions of multi‐view main‐scale representation learning and gated residual correction. Graphics processing unit (GPU) benchmarks showed that the dual‐window architecture incurred additional latency and memory relative to lightweight single‐window baselines. Overall, the results suggest that using long‐window information as a gated residual correction can improve performance in the intended one‐step (0.5 h) multivariate ocean wave nowcasting task while preserving the dominant role of recent observations.
ABSTRACT Multicore architectures have become the prevailing processor design for general‐purpose computing systems, enabling high performance across a wide range of commercial platforms. Technological advances have made it possible to integrate hundreds of cores on a single package, providing unprecedented computational capacity. In cloud data centers, efficiently exploiting this growing core count becomes crucial for improving resource efficiency, lowering operational costs, and boosting revenue. To achieve this, servers usually run multiple colocated cloud services alongside diverse workloads. However, fully utilizing all available cores while enforcing quality‐of‐service (QoS) constraints for critical services remains a major challenge. This paper introduces a novel OS‐level approach designed to maximize CPU usage in multicore servers through the opportunistic acceleration of elastic HPC workloads. These workloads are capable of dynamically adjusting the number of active worker threads/processes at runtime. Our proposal, referred to as EMC+, is an OS‐driven elasticity manager for container‐based environments that continuously estimates idle core cycles left by regular (inelastic) applications, and reallocates idle cores to elastic ones, even during short time intervals. The new EMC+ proposal builds on our previous work and broadens the scope of OS‐driven elasticity through several key contributions: the design of a new kernel‐level container‐management framework, efficient operation on large‐scale multicore platforms, the integration of QoS concerns into elasticity exploitation, and a comprehensive experimental analysis using diverse workloads that combine cloud services and HPC applications based on different parallel programming models. Experimental results show that EMC+ speeds up elastic workloads by up to 2.3 (1.5 on average), while having minimal impact on the performance and QoS of colocated inelastic applications. Compared to the unmodified Linux kernel, which does not exploit opportunistic elasticity, EMC+ boosts average system throughput by 19%, while degrading inelastic application performance by only 2.2% on average, and increasing the occurrence of observation intervals with QoS violations by 3.1% on average.
ABSTRACT The Log‐Structured Merge Tree (LSM‐tree) is widely adopted in key‐value (KV) storage systems because it transforms random writes into sequential writes. It has also become an important storage structure for supporting highly concurrent data access in high‐performance computing, cloud computing, and large‐scale data‐intensive applications. However, LSM‐tree still suffers from write amplification and periodic performance degradation, especially on emerging storage devices with strict sequential‐write constraints, such as Zoned Namespace Solid State Drives (ZNS SSDs). Meanwhile, Non‐Volatile Memory (NVM) provides new opportunities for optimizing the internal organization of LSM‐tree‐based storage systems. We observe that traditional SSTables tightly couple metadata with data blocks, causing frequent metadata updates to trigger unnecessary rewrites of large data regions and further exacerbating write amplification. To address this problem, we propose DS‐LSM‐tree, an LSM‐tree design based on SSTable decoupling. In DS‐LSM‐tree, large and low‐update‐frequency data blocks are sequentially stored on ZNS SSDs, while frequently updated fine‐grained metadata are migrated to NVM, thereby alleviating the mismatch between update granularity and storage device characteristics. Based on the decoupled SSTable design, we further develop a ZNS‐aware SSTable metadata structure and propose an NVM‐based logical merge strategy for SSTable data blocks, which reduces unnecessary physical data migration during compaction. We implement a prototype system named SPMS based on DS‐LSM‐tree and evaluate it using YCSB benchmarks. Experimental results show that, compared with ListDB, SPMS improves write throughput by up to 51.2%, increases read throughput by 5.1%, and reduces write amplification by 53.1%.
ABSTRACT The early detection of diseases in plants is essential to achieve maximum yield and minimum crop loss. This paper introduces an effective deep learning‐based pipeline to address the challenges, including a new FRCNN‐DA model for corn leaf disease classification and detection. The study consists of analyzing six models including FRCNN‐DA, OFDNN‐PDDC, CNN‐VGG, AHKM‐FCM, EfficientNet and ETL‐NET across four corn leaf type categories namely Northern leaf blight, Common rust, gray leaf spot and Healthy. The efficiency of the model was tested using performance metrics like Accuracy, Sensitivity, Specificity, F1‐Score, MAE, Detection Rate and AUC‐ROC. The FRCNN‐DA model achieved state‐of‐the‐art results with an accuracy of 97.85%, a sensitivity of 98.12% and an AUC score of 0.993 higher than all the benchmark models. Experimental results, combined with detection rate graph, MAE over 100 epochs and training loss graphs also show that the FRCNN‐DA converges rapidly in experimental learning, has high accuracy and is efficient for generalization. This framework provides a real‐world solution for utilizing AI to intelligently diagnose crop diseases in precision agriculture that could be extremely useful in decision‐making for crop health management.
ABSTRACT With the exponential growth in core density and workload complexity in modern chips, traditional two‐dimensional (2D) Network‐on‐Chip (NoC) architectures face critical limitations in scalability, thermal management, and energy efficiency. While three‐dimensional (3D) NoCs with wireless interconnects (WiNoC) offer improved vertical bandwidth and spatial integration, existing solutions such as DyAD, RC‐NoC, and CE‐NoC often rely on static routing and isolated flow control strategies that fail under dynamic traffic conditions, leading to congestion, energy inefficiency, and latency bottlenecks. Addressing these gaps, this study proposes CLARET (Cross‐Layer Adaptive Routing and Energy‐efficient Transport), a novel, decentralized communication framework that integrates dynamic flow control, cost‐based adaptive routing, and real‐time feedback across physical, MAC, and network layers. CLARET is implemented using a custom cycle‐accurate Python‐based 3D WiNoC simulator that incorporates both synthetic and trace‐driven traffic models. The framework dynamically adjusts packet injection rates, reroutes flows based on energy, congestion, and thermal metrics, and employs predictive traffic estimation using exponential smoothing. Experimental results show that CLARET reduces latency by up to 20%, lowers energy consumption by 14.4%, and enhances throughput by 23.5%, outperforming state‐of‐the‐art methods under varying traffic scenarios and workloads. The system maintains a packet delivery accuracy of 92%, marking a significant performance gain over existing protocols. These results demonstrate that CLARET is a scalable, energy‐aware communication model with real‐time adaptability—opening a promising avenue for intelligent, high‐performance NoC architectures. The proposed framework, with its unique cross‐layer synchronization and predictive control, offers a transformative step forward for next‐generation multicore design.
ABSTRACT The type of tobacco is closely related to the conditions and environment required for its processing and production. Efficiently and accurately identifying the different types of tobacco improves the accuracy of product quality control. Aiming at the current phenomenon of low efficiency of manual identification and screening of tobacco, poor classification effect leading to the difficulty of guaranteeing the quality of cigarettes, as well as the problem that the existing methods are easily interfered with by environmental factors with poor generalization ability in identifying the type of tobacco, and the accuracy is not high, a method of determining the type of tobacco components of the Golden Leaf based on the improvement of YOLOV9s was proposed. First, the original backbone network was replaced with FasterNet to reduce channel redundancy and model complexity while maintaining comparable detection performance. This lightweight backbone provides a computationally efficient foundation for subsequent feature enhancement modules. Second, in the feature extraction part, variable kernel convolution AKConv was used to replace the original traditional convolution, which utilizes a flexible convolution mechanism to keep the model lightweight and enhance the accuracy of target information feature extraction simultaneously. Finally, the downsampled convolutional AConv of the neck was replaced with HWD (Haar wavelet downsampling) to expand the receptive field and reduce the loss of important spatial information. It was experimentally verified that compared to the original YOLOV9s algorithm, the improved algorithm improved recall and mAP by 2.33% and 1.1%, respectively, and reduced the number of references and memory footprint by 1.16 × 10 6 and 2.22 MB, respectively. YOLOV9s‐RepNCSPELAN4_AKConv‐HWD was 1.92%, 2.82%, 2.67%, and 5.76% higher in mAP than other mainstream models. The results showed that the algorithm struck a good balance between detection accuracy and lightweight, providing a reliable reference with high accuracy and real‐time performance for deployment on local devices.
