Concurrency and Computation-Practice & Experience最新文献

Surface defect detection is critical in industrial applications, especially for multi-scale small object detection under complex backgrounds, low resolution, and real-time constraints. Traditional YOLO-based detectors face three key limitations: single-path backbones struggle to balance fine-grained details and high-level semantics simultaneously, conventional feature pyramids cause small-object feature dilution through simple additive fusion, and detection heads lack explicit quality estimation for diverse defect morphologies. To address these challenges, we propose RFG-YOLO, a lightweight detection network. The proposed method introduces three key innovations: DPFNet employs dual-path parallel processing with PSHGNetv2 and YOLOv11 backbone to extract fine-grained details and semantic context simultaneously, eliminating the single-path bottleneck; MSWFPN utilizes weighted fusion modules including MFM, SCFM, and CSP-SMSFB to prevent small-object feature dilution through adaptive multi-scale integration; RQCD incorporates location quality estimation with RepConv to refine bounding box predictions for diverse defect shapes, improving localization precision across scales. Experiments on NEU-DET, GEAR-DET, and GC10-DET demonstrate that RFG-YOLO achieves mAP@50 improvements of 4.1%, 5.77%, and 4.1%, respectively, while reducing model parameters by 60.5% and computational cost by 50.8%. The model achieves a runtime speed of 156.0 frames per second on a GPU. These results validate that RFG-YOLO strikes an optimal balance between detection accuracy and computational efficiency, rendering it highly suitable for real-time defect detection in industrial settings.

The factorized sparse approximate inverse (FSAI) preconditioner has been proven to be effective in accelerating the convergence of iterative methods. Due to the high cost of constructing the FSAI preconditioner, accelerating it on graphics processing unit (GPU) has attracted considerable attention. However, despite the development of some existing FSAI preconditioning algorithms on GPU, their performance will significantly decrease when they encounter matrix types that are not suitable for them. This motivates us to investigate how to design an effective FSAI preconditioning algorithm on GPU. In this paper, we propose an adaptive FSAI preconditioning algorithm on GPU, called GFSAI-Adaptive, to address the above problem. In GFSAI-Adaptive, first, two adaptive thread allocation strategies are proposed for two special types of SPD matrices to ensure that the allocated threads can be fully utilized. Second, based on the proposed two thread allocation strategies, two FSAI kernels, called GFSAII and GFSAIII, are presented. Third, we construct a new graph convolutional network, and thus propose a search engine to select the optimal kernel from GFSAII and GFSAIII for matrices that do not belong to two special types based on it. Experimental results show that our proposed GFSAI-Adaptive is effective and outperforms a popular preconditioning algorithm in the public CUSPARSE library and a recent parallel static FSAI preconditioning algorithm on GPU.

The management of agricultural pests is crucial for the security of the food supply. The Brown Plant Hopper (BPH) harms rice fields and reduces productivity, which is a major concern. Accurate forecasting and early detection of BPH population dynamics based on environmental parameters are essential for the quick and efficient application of pest control strategies. This study investigates the use of machine learning methods like bagging, boosting, voting, and ensemble regression techniques to predict the population of BPH. The proposed ensemble stacking regression model uses Extreme gradient boosting (XGBoost), Random Forest (RF), and Extremely Randomized Trees (ERT) as base models with a meta-model Decision Tree (DT), and it is utilized to estimate the pest population. Several existing models are compared with the proposed model in terms of R-squared, Adjusted R-squared, RMSE as well as MSE. The outcomes show that these machine learning models are capable of accurately and successfully predicting the BPH population, which makes the proposed model an important tool for pest population prediction in rice farming.

Knee osteoarthritis (KOA) grading from x-ray images is important for supporting effective treatment planning. Yet it remains difficult due to the disease's complex presentation. Subtle anatomical changes add to the challenge. The subjectivity of manual evaluation further complicates the process. Such challenges highlight the importance of automated, objective, and reproducible computer-aided systems capable of leveraging complementary sources of information. In line with this, a joint fusion framework was developed to integrate optimized traditional image features with deep learning representations obtained from multiple pre-trained convolutional neural network models. Traditional features, including morphological, statistical, texture-based, and other clinically relevant descriptors, provide interpretable insights into bone structure and tissue characteristics. In parallel, deep features capture intricate spatial patterns and semantic details beyond the reach of manual modeling. For improved discrimination and efficiency, analysis of variance and linear discriminant analysis were used for selecting traditional features, while principal component analysis was applied to deep features to retain 85% variance. The two feature sets were combined using a Joint Fusion Type II approach, and class imbalance was mitigated through synthetic minority oversampling. A neural network was trained to capture interdependencies between these features. Experimental results on a benchmark KOA dataset indicated that fusion with VGG16 deep features achieved 85.39% accuracy, outperforming individual feature-based approaches. The framework maintained relatively high accuracy across all five KOA grades, including borderline cases, indicating its potential for consistent and clinically relevant KOA grading.

The proliferation of Open Radio Access Network (O-RAN) architectures in 5G/6G networks introduces unprecedented cybersecurity challenges. Strategic base station deployment constitutes a fundamental determinant of network security posture and cyberattack resilience. In this paper, a novel cybersecurity-driven deployment strategy for resilient base station positioning using an intelligent Resilient Ant Colony Optimization (iResACO) algorithm. The algorithm integrates security considerations directly into deployment optimization, employing bio-inspired collective intelligence to discover patterns that balance coverage efficiency with attack resilience. Through extensive simulations in a 3.6 km $��\times �$ 3.6 km urban environment in Riyadh, Saudi Arabia, experimental results demonstrate superior performance achieving 92.04% overall effectiveness with 96.0% coverage probability and 100% critical infrastructure protection. Under various cyberattack scenarios ranging from random to coordinated sophisticated attacks, the algorithm maintains coverage above 87% while preserving complete protection of critical facilities. The proposed approach provides a practical framework for deploying secure, resilient 5G/6G networks capable of withstanding evolving cyber threats while ensuring uninterrupted service to essential infrastructure.

As model parameters increase exponentially, distributed training has become essential for advancing modern deep neural networks. Megatron-LM, an efficient distributed training framework developed by NVIDIA, enables the training of trillion-parameter models on thousands of GPUs by integrating tensor, pipeline, and data parallelism. Its computational efficiency has established it as a foundational tool for training large-scale models. Rapid identification of optimal parallel configurations for specific GPU clusters is critical for maximizing computational resource utilization, with training time prediction serving as a key evaluation metric. The high cost and limited availability of high-performance GPUs, particularly those based on NVIDIA architectures, have made the construction of large-scale heterogeneous clusters a practical solution to resource and cost constraints. However, existing prediction methods do not reliably or efficiently account for the computational and communication complexities inherent in heterogeneous GPU clusters. To address this gap, HATP (Heterogeneous-Aware Time Predictor) is introduced as a novel performance prediction method specifically designed for heterogeneous GPU clusters. For any given parallel configuration, HATP rapidly and accurately simulates execution times to inform the optimization of parallel strategies. To address communication differences among heterogeneous GPUs, comprehensive experimental analyses are conducted and analytical expressions are derived to characterize the communication frequency patterns in Megatron-LM's parallel strategies. This work presents the first systematic quantification of communication operations within Megatron-LM framework, ensuring that performance predictions remain highly accurate even in complex, heterogeneous environments. Furthermore, to account for computational differences among heterogeneous GPUs, a layer-level computational performance acquisition scheme is proposed to reduce the impact of fine-grained operator overlap and additional memory operations. Experimental results demonstrate that HATP achieves an average prediction accuracy of 97.41% in isomorphic environments, surpassing the current state-of-the-art method, ACEso. HATP also attains an average accuracy of 96.04% in heterogeneous data parallel and pipeline parallel configurations, representing the first extension of training time prediction capabilities to heterogeneous environments.

The increasing application of unmanned aerial vehicles (UAVs) in diverse domains demands highly robust and autonomous path-planning algorithms capable of navigating complex and dynamic environments. To address the multifaceted challenges posed by obstacle avoidance, energy constraints, and environmental uncertainty, this work proposes an ensemble of learning strategies for the optimal path planning of UAVs. We introduce a modular particle swarm optimization and differential evolution (PSO-DE) ensemble framework and systematically investigate the impact of multiple learning and adaptation strategies, such as chaotic parameter adaptation, opposition-based learning (OBL), and a range of DE mutation schemes, to enhance the optimization process. We perform extensive experimentation across 16 carefully designed scenarios with varying complexity against ten competitive algorithms. We demonstrate that the integration of the PSO-DE hybrid with the opposition-based learning (OBLPSODE) achieves faster convergence while maintaining superior solution quality across all scenarios. The proposed OBLPSODE algorithm substantially outperforms other hybrid variants in both computational efficiency and path optimality, particularly excelling in cluttered environments where traditional algorithms often converge prematurely. Beyond algorithmic contributions, this work provides critical complexity analysis identifying obstacle-checking operations as the primary computational bottleneck in UAV path planning. The findings offer practical guidance for deploying UAVs in real-world applications and establish transferable design principles for developing adaptive meta-heuristics in complex optimization domains.

Vessel re-identification (ReID) plays a critical role in maritime surveillance by matching vessels across different camera views. Compared with person or vehicle ReID, vessel ReID faces unique challenges due to subtle interclass differences and large intraclass variations caused by viewpoint changes. These issues are further exacerbated by the highly similar appearances of vessels and the lack of fine-grained identity cues commonly found in other ReID tasks. To address these challenges, we propose a spatial-channel fusion network (SCF-Net), a dual-branch deep framework that integrates a spatial-channel fusion (SCF) module and a feature refinement and alignment (FRA) module. The SCF module captures interdependent relationships between spatial and channel dimensions, enabling the network to emphasize discriminative regions while suppressing irrelevant background information. The FRA module refines high-dimensional embeddings into a compact representation and enforces intraclass similarity via a learnable multilayer perceptron (MLP) and a supervised mean squared error (MSE) loss. By jointly optimizing the two branches and the FRA output, SCF-Net effectively learns both interclass discrimination and intraclass compactness. Extensive experiments demonstrate that SCF-Net achieves competitive performance on public vessel ReID benchmarks, highlighting its effectiveness in handling subtle interclass differences and large intraclass variations.

Tunnel face instability prediction represents a critical technical challenge in underground engineering, particularly during tunnel boring machine (TBM) excavation under complex geological conditions. This study proposes the Multi-modal Gaussian Cross-Attention Fusion (MGCAF) algorithm, which integrates physics-constrained Gaussian processes with cross-attention mechanisms to achieve intelligent tunnel collapse prediction. The MGCAF framework reconstructs the traditional prediction paradigm by treating earth pressure balance chamber pressure as the primary prediction target rather than an input parameter, while incorporating first-principles constraints of TBM cutting mechanisms into kernel function design. The algorithm employs a dual-pathway architecture that fuses TBM operational parameters through temporal modeling, processes geological radar images via deep feature extraction, and achieves cross-modal information fusion through physics-constrained cross-attention mechanisms. Dynamic kernel optimization enables real-time adaptive parameter adjustment through multi-source gradient feedback. Validation results based on the Yinsong Water Diversion Tunnel project (20 km length, 9 collapse events) demonstrate that the algorithm achieves high-precision prediction with R² = 0.8330, successfully predicting major collapse locations with approximately 20-m accuracy. Comparative analysis against baseline methods (Transformer, Gaussian Process, Random Forest, XGBoost) indicates that MGCAF exhibits superior performance in engineering reliability (0.95) and ROC-AUC (0.765) metrics. Generalization testing on the 2025 Los Angeles Wilmington Sewage Outfall Tunnel confirms the algorithm's cross-domain applicability. Ablation experiments reveal that the cross-attention mechanism serves as the primary performance driver, while uncertainty quantification provides interpretable risk assessment for TBM operations in heterogeneous geological environments.

The growing demand for data sharing in domains such as health care highlights limitations in existing solutions, including low search efficiency, coarse-grained access control, and heavy decryption overhead on users. To address these challenges, this paper proposes a hierarchical searchable attribute-based encryption scheme with outsourced decryption for blockchain-based data sharing (OD-H-SABE). OD-H-SABE introduces a hierarchical attribute structure alongside an outsourced decryption mechanism that offloads computationally intensive bilinear operations to the cloud server. Consequently, users only need to perform a single lightweight operation to complete decryption, significantly alleviating the computational burden. Furthermore, the scheme integrates searchable encryption with multi-keyword aggregate hashing, enabling efficient search with constant complexity regardless of the number of keywords. Leveraging the transparency and immutability of blockchain, smart contracts verify the integrity of results returned from the cloud server, ensuring data security and trustworthiness throughout the sharing process. Theoretical and experimental analyses demonstrate that OD-H-SABE achieves notable advantages over traditional schemes in terms of security, search efficiency, and computational overhead. For example, compared to MKS-VABE and BEM-ABSE, OD-H-SABE reduces encryption and user-side decryption overhead by approximately 20% and 42%, respectively. This makes it a practical and lightweight solution for constructing secure and efficient blockchain-based data-sharing platforms.