Object detection is a critical component in the development of autonomous driving technology and has demonstrated significant growth potential. To address the limitations of current techniques, this paper presents an improved object detection method for autonomous driving based on a detection transformer (DETR). First, we introduce a multi-scale feature and location information extraction method, which solves the inadequacy of the model for multi-scale object localization and detection. In addition, we developed a transformer encoder based on the group axial attention mechanism. This allows for efficient attention range control in the horizontal and vertical directions while reducing computation, ultimately enhancing the inference speed. Furthermore, we propose a novel dynamic hyperparameter tuning training method based on Pareto efficiency, which coordinates the training state of the loss functions through dynamic weights, overcoming issues associated with manually setting fixed weights and enhancing model convergence speed and accuracy. Experimental results demonstrate that the proposed method surpasses others, with improvements of 3.3%, 4.5%, and 3% in average precision on the COCO, PASCAL VOC, and KITTI datasets, respectively, and an 84% increase in FPS.
Abdulrahman M. Abdulghani, Azizol Abdullah, A. R. Rahiman
et al.
Modern Software-Defined Wide Area Networks (SD-WANs) require adaptive controller placement addressing multi-objective optimization where latency minimization, load balancing, and fault tolerance must be simultaneously optimized. Traditional static approaches fail under dynamic network conditions with evolving traffic patterns and topology changes. This paper presents a novel hybrid framework integrating Gaussian Mixture Model (GMM) clustering with Multi-Agent Reinforcement Learning (MARL) for dynamic controller placement. The approach leverages probabilistic clustering for intelligent MARL initialization, reducing exploration requirements. Centralized Training with Decentralized Execution (CTDE) enables distributed optimization through cooperative agents. Experimental evaluation using real-world topologies demonstrates a noticeable reduction in the latency, improvement in network balance, and significant computational efficiency versus existing methods. Dynamic adaptation experiments confirm superior scalability during network changes. The hybrid architecture achieves linear scalability through problem decomposition while maintaining real-time responsiveness, establishing practical viability.
Ivana Valentina Lemmuela, Mewati Ayub, Oscar Karnalim
Background: Communication is important for everyone, including individuals with hearing and speech impairments. For this demographic, sign language is widely used as the primary medium of communication with others who share similar conditions or with hearing individuals who understand sign language. However, communication difficulties arise when individuals with these impairments attempt to interact with those who do not understand sign language.
Objective: This research aims to develop models capable of recognizing sign language movements in Bahasa and converting the detected gesture into corresponding words, with a focus on vocabularies related to religious activities. Specifically, the research examined dynamic sign language in Bahasa, which comprised gestures requiring motion for proper demonstration.
Methods: In accordance with the research objective, sign language recognition model was developed using MediaPipe-assisted extraction process. Recognition of dynamic sign language in Bahasa was achieved through the application of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) methods.
Results: Sign language recognition model developed using bidirectional LSTM showed the best result with a testing accuracy of 100%. However, the best result for the CNN alone was 86.67 %. The integration of CNN and LSTM was observed to improve performance than CNN alone, with the best CNN-LSTM model achieving an accuracy of 95.24%.
Conclusion: The bidirectional LSTM model outperformed the unidirectional LSTM by capturing richer temporal information, with a specific consideration of both past and future time steps. Based on the observations made, CNN alone could not match the effectiveness of the Bidirectional LSTM, but a combination of CNN with LSTM produced better results. It is also important to state that normalized landmark data was found to significantly improve accuracy. Accuracy within this context was also influenced by shot type variability and specific landmark coordinates. Furthermore, the dataset containing straight-shot videos with x and y coordinates provided more accurate results, dissimilar to those comprised of videos with shot variation, which typically require x, y, and z coordinates for optimal accuracy.
Keywords: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), MediaPipe, Sign Language
Federated learning (FL) enables collaborative model training across multiple institutions without the sharing of raw patient data, making it particularly suitable for smart healthcare applications. However, recent studies revealed that merely sharing gradients provides a false sense of security, as private information can still be inferred through gradient inversion attacks (GIAs). While differential privacy (DP) provides provable privacy guarantees, traditional DP methods apply uniform protection, leading to excessive protection for low-sensitivity data and insufficient protection for high-sensitivity data, which degrades model performance and increases privacy risks. This paper proposes a new privacy notion, sensitivity-aware differential privacy, to better balance model performance and privacy protection. Our idea is that the sensitivity of each data sample can be objectively measured using real-world attacks. To implement this new notion, we develop the corresponding defense mechanism that adjusts privacy protection levels based on the variation in the privacy leakage risks of gradient inversion attacks. Furthermore, the method extends naturally to multi-attack scenarios. Extensive experiments on real-world medical imaging datasets demonstrate that, under equivalent privacy risk, our method achieves an average performance improvement of 13.5% over state-of-the-art methods.
The development of society is accompanied by an increase in the complexity of management objects and management mechanisms. To counteract the growth of complexity, new management models and methods should be introduced. New methods include semasiological management which uses a model approach and induction principle. It borrows the ideas of semasiology from linguistics and forms management decisions on the basis of application of information management units. Despite the fact that this complicates the preliminary process of preparing for management, it also gives an advantage in the comparability of different management decisions and technologies. Semasiological management allows, when reconfiguring management, not to create management models anew, but to modernise them by replacing management information units or forming new combinations of these units. Semasiological management is related to onomasiological information modeling and requires its use. In addition, it can be used in automated management, smart management, and digital twin management. Semasiological management requires special organisation and specific training, such as a special management language. The research proposes a variant of semasiological management which is based on the application of the theory of information units.
Hematopoietic stem cell transplantation (HSCT) is a crucial treatment option for hematological disorders. This study aims to design and develop a blockchain-based system tailored explicitly for HSCT typing and evaluate its performance and capabilities. The system, built on Hyperledger Fabric 2.4.6 and integrated with LevelDB and MySQL databases, implements features such as data on-chaining, access control, information retrieval, and matching through chaincode. It ensures accurate and efficient human leukocyte antigen (HLA) typing while protecting patient and hospital data security. Tests demonstrate the system’s ability to safeguard data security and reliability effectively. Performance metrics, including block size, CPU utilization, network throughput, and response latency, were evaluated on Ubuntu 20.04.2 operating systems using VMware Workstation 16 Pro and Docker containerization. By leveraging blockchain’s traceability and immutability, the system achieves data security, reliability, and verifiability, enabling faster and more accurate typing and improved healthcare efficiency. This innovative approach optimizes user experience, maintains data integrity, and ensures privacy. Future integration with existing databases could enable secure data sharing among healthcare institutions, allowing rapid verification of data ownership and unified authentication. This study contributes a significant advancement in hematologic disease treatment and research, offering practical implications for improving HSCT processes.
The current article investigates the numerical study of the micropolar nanofluid flow through a 3D rotating surface. This communication may manipulate for the aim such as the delivery of the drug, cooling of electronic chips, nanoscience and the fields of nanotechnology. The impact of heat source/sink is employed. Brownian motion and thermophoresis aspects are discussed. The rotating sheet with the impacts of Darcy-Forchheimer law is also scrutinized. Furthermore, the influence of activation energy is analyzed in the current article. The numerical analysis is simplified with the help of befitted resemblance transformations. The succor of the shooting algorithm with built-in solver bvp4c MATLAB software is used for the numerical solution of nonlinear transformed equations. The consequences of different physical factors on the physical engineering quantities and the subjective fields were examined and presented. According to outcomes, it can be analyzed that the flow profile declined with the rotational parameter. It is observed that angular velocity diminishes via a larger porosity parameter. Furthermore, the temperature gradient is declined via a larger magnitude of the Prandtl number. The heat transfer is enhanced in the occurrence of Brownian motion. The activations energy parameter causes an increment in the volumetric concentration field. Moreover, the local Nusselt number is reduced via a greater estimation of the porosity parameter.
Muhammad Irfan, Zohaib Mushtaq, Nabeel Ahmed Khan
et al.
Machine learning (ML) based bearing fault detection is an emerging application of Artificial Intelligence (AI) that has proven its utility in effectively classifying various faults for timely measures. There are myriad studies dedicated to the effective classification of bearing faults under different conditions and experimental settings. In this study, we proposed a weighted voting ensemble (WVE) of three low-computation custom-designed convolutional neural networks (CNNs) to classify bearing faults at 48 KHz. Some of the recent studies have exploited 1-d time-series signals and time-frequency based 2-d transformations for bearing fault classification. However, 1-d signals lack contextual information and higher-dimensional interpretations whereas time-frequency based transformations provide a more appropriate, visually perceivable and explainable representation of the time and frequency changes. Therefore in this study, a scalogram based representation of the signals is leveraged for classification using the CNN. Furthermore, the class imbalance is a significant challenge that affects the modelling behavior and possibly create biases. This study provides a novel density and distance hybrid over-sampling approach namely Density-Aware SMOTE(DA-SMOTE) built upon the SMOTE methodology for a more refined representation of synthetic samples within the minority class distribution. The experimentation procedures were carried out before and after the oversampling and it was observed that the balanced dataset acquired much better accuracy then the imbalanced dataset. This is evident by the fact that the highest validation accuracy for the proposed ensemble method (WVCNN) reached at 0-HP and 1-HP reached 99.28% and 99.13% while for the over-sampled dataset the accuracy soared to 99.71% and 99.87% for 0 and 1-HP respectively. The performance was evaluated for other metrics apart from the accuracy to assess the model’s performance in terms of chance occurrences and the class wise performance.
The target simulation of airplanes is an important research topic. It is particularly important to find the right balance between high performance and low cost. In order to balance the contradictions between realistic target simulations and controllable costs, the scientific formulation of the performance parameters of target simulation is the key to achieving high performance. This paper proposes an intelligent simulation technology based on RCS imaging simulation through the combination of 60° variation corner reflector and a Luneberg lens reflector. It is designed to simulate several important RCS characteristics of the aircraft. At the same time, the different RCS images are automatically shifted to the corresponding gear position to achieve the purpose of simulation, and the price is low and the performance is good. It can be used for the training of radar target searching.