AUTOMATIC CONTROL AND COMPUTER SCIENCES最新文献

Satellite image data classification is a crucial task for many applications such as urban planning, environmental monitoring, national border security, etc. In the era of artificial intelligence, neural network approaches for satellite image classification have shown good results. Transformer based approaches have completely transformed the artificial intelligence methods in the last five years. Initially the transformer approaches have been proposed for text processing. For computer vision problems, Vision Transformer has been proposed in year 2020, which is utilized for many research applications in areas of healthcare, satellite imagery, defence etc. Transformer based models have demonstrated that the attention mechanism plays a crucial role. In this paper, a specialized attention mechanism approach focused on spatial, spectral, and temporal features of satellite image combined with a vision transformer, is proposed. The proposed architecture is known as Context-Aware Vision transformer (CAViT) for satellite image classification. We applied the proposed model on publicly available three satellite image scene datasets i.e., University of California Merced (UCM) with 21 classes, Aerial Image Dataset (AID) with 30 classes, and Remote Image Scene Classification dataset of Northwestern Polytechnical University (NWPU-RESISC45) with 45 classes. We used performance metrics parameters as accuracy, recall, precision, F1-score, and confusion matrix to evaluate the model’s performance with different datasets. This model achieved an overall accuracy of 99.33% for UCM, 97.71% for AID, and 95.63% for NWPU-RESISC45 dataset. The model shows competitive results against other deep learning models. Our research paper revealed CAViT proficiency in the satellite image classification applications ranging from environment monitoring to urban planning.

This paper presents a dynamic lung infection detection system utilizing advanced deep learning algorithms to analyze 3D CT images. The primary goal is to improve the accuracy and reliability of detecting lung infections, which is critical for providing timely medical intervention and enhancing patient outcomes. We investigated and compared two sophisticated models: a 3-dimensional Convolutional Neural Network (3D CNN) and a Residual Network (3D ResNet 101). These models were implemented in Python and rigorously evaluated on the MosMed dataset. The evaluation included key performance metrics such as accuracy, precision, recall, and F1 score, to assess their effectiveness in diagnosing lung infections from high-resolution 3D CT scans. The 3D CNN model demonstrated exceptional performance, achieving an accuracy of 99.60%, precision of 99.73%, recall of 96.80%, and an F1 score of 97.25%. In comparison, the 3D ResNet 101 model reached a maximum accuracy of 97.30%, precision of 99.25%, recall of 95.28%, and an F1 score of 96.32%. These results underscore the 3D CNN model’s superior performance in detecting lung infections. The study highlights the effectiveness of the 3D CNN model in lung infection detection, surpassing the 3D ResNet 101 model in several key performance metrics. This demonstrates that the integration of cutting-edge deep learning techniques with high-resolution 3D CT imaging offers significant advancements in diagnostic accuracy. The findings suggest that the 3D CNN model holds promise for enhancing diagnostic procedures and improving patient care in clinical settings. Future work will focus on further optimizing these models and exploring their applicability to other medical imaging tasks.

The Internet of Things records a large amount of user information in the process of operation, which causes users to face the security risk of information leakage. The study presents blockchain technology, evaluates attack detection and node authentication techniques, and suggests network attack detection and node authentication techniques based on blockchain technology to ensure the information security of wireless sensor networks. The results revealed that all the nodes are infected in less than 20 time slots when the network nodes are 300 and after 50 time slots when the network nodes are 75, the worm attack completes the attack on all the nodes. The infection rate of the worm attack was increased gradually as the network nodes increases. The F1 value of the proposed scheme of the study increases with the advancement of time slots and can eventually increase to around 0.88. The research has developed a wireless sensor network node authentication in attack detection method that can effectively distinguish between normal and abnormal activity in the network, thereby realizing the wireless sensor network’s attack warning function. The research-designed sensor network node authentication technology and attack detection technology strengthens the confidentiality level of user information and promotes the popularization and application of wireless sensor networks.

Violence refers to the intentional use of physical force, power, or coercion against oneself, another person, or a group, resulting in harm, injury, or deprivation. Public safety represents a multifaceted and intricate challenge, demanding swift identification and prevention of violent incidents. Violence detection is a critical aspect of maintaining public safety and security in various domains, including surveillance, law enforcement, and online content moderation. This paper proposes a holistic approach to enhancing public safety through the advanced analysis of Closed-Circuit Television (CCTV) footage. The envisioned framework presents a system acting as an electronic guardian, tirelessly surveilling environments and promptly identifying potential violence through digital technologies. By integrating Internet of Things (IoT) devices and alarm sensors, this system operates as a vigilant observer, capable of detecting signs of aggression and stress even in dynamic or chaotic settings. Through the application of deep learning techniques, the system endeavors to replicate human observation, issuing alarms swiftly to alert security personnel. By harnessing the power of IoT and deep learning, this approach represents a paradigm shift in public safety, offering a proactive and responsive solution to the challenges posed by emerging threats. Ultimately, the proposed framework establishes a symbiotic relationship between technological innovation and public safety, paving the way for safer and more secure communities.

An adaptive sliding mode controller (ASMC) is presented that can control plants of the n-th order with completely unknown parameters, while using a varying boundary layer thickness for complete chatter prevention. Adaptation laws are derived for parameter estimation and varying boundary layer thickness adaptation. A Lyapunov stability analysis shows that the parameter adaptation laws drive the combined error for the sliding surface to zero while rejecting bounded disturbances. The boundary layer adaptation law automatically grows the boundary layer thickness in the presence of disturbances and reduces it when the disturbances are not present to always keep the sliding surface error trajectory within the boundary layer, which guarantees chatter prevention. Simulations show the superior performance of the proposed ASMC compared to a non-adaptive sliding mode controller with a fixed boundary layer thickness.

In response to the current problem of unsatisfactory enhancement effect and low image processing efficiency of underwater images using dark channels, an improved algorithm based on dark channel priors is proposed to process underwater images. The method based on median filtering is used to obtain the atmospheric transmittance t, which improves the efficiency of atmospheric transmittance t acquisition under the premise of ensuring the defogging effect. The quadtree image segmentation method is used to obtain the global atmospheric light A, which improves the image processing effect. Aiming at the problem of insufficient saturation of the image and low contrast of local details, the image is transferred to HSV space and enhanced by adaptive saturation adjustment and Gamma correction respectively. Four images are selected for experiments and analysis. The results show that compared with the original algorithm, the improved algorithm increases efficiency by about 34% while ensuring the defogging effect. Moreover, the improved algorithm restores more detailed information in the image and removes the fog from the image effectively.

To improve the accuracy of traffic sign recognition in complex backgrounds and extreme conditions, an improved YOLO network deep learning method is proposed. This method achieves cross scale connection and fast normalization fusion of multiple features through label smoothing and loss function improvement, and introduces a mixed attention mechanism to enhance the robustness of the recognition process. The experimental results show that our method can effectively cope with the impact of complex backgrounds and extreme conditions on the recognition process, and the accuracy of traffic sign recognition is significantly higher than the three methods of CNN, RNN, and YOLO.

For the multi-objective optimization problem of energy saving, comfort, punctuality and on-time performance in the process of Automatic Train Operation (ATO) of high-speed trains, a solution algorithm based on particle swarm algorithm with adaptive hybrid strategy is proposed. Firstly, for the inaccuracy of the force analysis of single-mass point modeling of high-speed train, a rigid multi-mass point model of high-speed train is established; secondly, using the dynamics model of high-speed trains and the safety of train operation as constraints, the affiliation function is used to establish a multi-objective optimization model of high-speed train ATO, when dealing with the constraints, the high-speed train stopping error and the line speed limit are used as penalty items to construct a suitable penalty function to be added to the objective function, which constitutes the fitness function used in this paper; finally, in order to solve the shortcomings of the particle swarm optimization algorithm that is easy to converge and easy to fall into the local optimum, the adaptive learning mixed strategy particle swarm optimization algorithm is proposed. Experimental validation is carried out by selecting real routes and high-speed trains to verify the effectiveness of the method proposed in the paper in reducing the energy consumption of high-speed train operation, improving comfort, and arriving at the destination on time and on schedule.

The Centroid Localization Algorithm (CLA) is a commonly used technique in wireless sensor networks (WSN) to detect the location of target nodes. Nonetheless, the localization errors associated with CLA are typically substantial, which can reduce its effectiveness in real-life WSN applications. To address this limitation, while achieving the basic localization goal of accurately identifying unknown nodes in the WSN, this paper proposes a new localization approach, namely, SLnA-CLA, by integrating the CLA and the sea lion optimization algorithm (SLnA), which is a bioinspired technique based on sea lion social behavior. We compare the performance of our proposed SLnA-CLA algorithm with the basic CLA and SLnA for nodes localization algorithms. In this study, we make sure to evaluate the three algorithms, SLnA-CLA, CLA, and SLnA, using the same deployment of anchor and target nodes. This way, we confirm that any performance discrepancies are attributable to the algorithms, not to any biases introduced by the various network topologies. The results demonstrate that the proposed algorithm effectively reduces localization error by up to 98.7% when compared to CLA, albeit with a longer computation time, and outperforms SLnA in both accuracy and computation time.

The high-precision time-frequency is a crucial strategic resource for any country, serving as the bedrock for national defense construction and economic operations. Global Navigation Satellite Systems (GNSS) provides timing service through satellite, which has the characteristics of all-weather, high-precision and wide coverage. It has become an indispensable means of timing service. However, GNSS signal can be disrupted under certain conditions, leading to service interruptions and preventing users from obtaining accurate time information. This paper introduces a time-frequency broadcasting system utilizing the BeiDou Navigation Satellite System (BDS) signal system in scenarios when GNSS is unavailable. By acquiring and calculating the BDS ephemeris data, the system calculates the BDS coordinate information according to the local time in real-time, simulates the satellite trajectory, adjusts the time of broadcasting according to the local position input, and generates the corrected satellite navigation message information. The time-frequency broadcasting system can provide real-time timing service for BDS navigation terminals within the coverage area. The timing accuracy of the system is better than 6.3 nanoseconds, the project implementation is feasible, and the application range is wide. In the case of no GNSS, the system can also provide emergency timing service for BDS navigation terminals within the coverage area.