Relation extraction is a crucial task in natural language processing, aimed at identifying entities and their relationships from text, and holds particular significance in the cryptography domain due to its involvement with complex multi-entity interactions. This paper addresses the challenges of relation extraction in cryptography, such as intricate terminology, implicit relationships, data scarcity, and the demand for high precision, by proposing a method based on retrieval augmentation and chain-of-thought optimization. By integrating chain-of-thought prompt optimization, example retrieval enhancement, and dependency analysis techniques, this approach significantly improves extraction performance in low-resource environments. Additionally, a high-quality cryptography-domain dataset containing 24 entity relationship categories is constructed, providing robust support for related research.

Tailored to the unique characteristics of the cryptography domain, this paper first designs a chain-of-thought-based prompt template with a self-checking mechanism to guide the model through step-by-step reasoning, enhancing the accuracy of extracting complex relationships. Second, it employs S-BERT embeddings along with a weighted ranking strategy using cosine similarity and the DICE coefficient to dynamically retrieve and inject the most relevant examples into the prompt template, improving the model’s adaptability in data-scarce scenarios. Then, a rule-based retrieval method precisely injects necessary background knowledge. Finally, StanfordCoreNLP is utilized to extract dependency relationships as structured knowledge, further optimizing the model’s understanding of implicit relationships.

Experimental results demonstrate that this method, implemented on the ChatGLM4-Plus model, achieves an F1-score of 70.62%, significantly outperforming traditional methods and unoptimized large language model baselines. Ablation studies confirm the effectiveness of each module and their synergistic contributions, underscoring the method’s value in enhancing the precision and robustness of relation extraction in the cryptography domain. This approach not only provides an efficient tool for knowledge graph construction and information retrieval in cryptography but also offers a replicable optimization strategy for relation extraction tasks in other highly specialized domains.et

As a key component of intelligent transportation systems, Vehicular Ad Hoc Network (VANET) has been widely used to improve traffic congestion, optimize driving paths, improve driving safety, and provide diverse entertainment services. However, due to its open communication architecture, the Internet of Vehicles faces many security challenges in message transmission and user identity protection. First, when vehicles communicate with other vehicles (Vehicle-to-Vehicle, V2V) or infrastructure (Vehicle-to-Infrastructure, V2I), information is vulnerable to security threats such as forgery, eavesdropping, message replay, and denial of service attacks. This may lead to driver misjudgment and safety accidents. Secondly, the Internet of Vehicles also faces problems in user privacy protection, such as location privacy leaks and identity information abuse. Compared with existing PKI and identity-based schemes, the certificateless architecture proposed in this paper has significant advantages and benefits in the following aspects: efficiency, fault tolerance, privacy protection, scalability, and security. Through security analysis and performance evaluation, the results show that this solution is superior to existing methods in terms of computing efficiency, storage requirements, communication overhead, etc., and on the basis of ensuring privacy protection and data integrity, it achieves high-dynamic environment Secure authentication and efficient communication.

Path planning is a critical problem in robotics, aimed at generating efficient and smooth trajectories for robots. However, traditional algorithms like A* often suffer from low search efficiency, redundant nodes, and discontinuities in velocity and acceleration at path corners. To address these issues, this paper presents a path planning framework that combines the Optimized Bidirectional A* algorithm, the Critical Node Retention algorithm, and the Minimum Snap algorithm. The framework enhances the path planning process in three stages: First, the Bidirectional A* search is optimized by utilizing a weighted heuristic function to improve search efficiency and reduce unnecessary node expansions. Second, the Critical Node Retention algorithm refines the path by retaining only critical nodes, reducing redundancy and shortening the path length. Finally, the Minimum Snap algorithm smooths the path, ensuring continuous velocity and acceleration, thereby generating safer and smoother trajectories. Simulation experiments on grid maps of various sizes demonstrate that the proposed framework effectively reduces path length, improves search efficiency, and generates smooth, safe trajectories, providing an efficient solution to path planning.

With the continuous development of deep learning in recent years, the field of image aesthetic caption has gradually become a popular research direction, which significantly impacts various applications, such as advanced semantic understanding of images and the promotion of artistic images. Image aesthetic caption generation is a cutting-edge direction for integrating computer vision and natural language processing. Unlike traditional image caption that focuses on outputting basic facts of images, the core goal of image aesthetic caption research is to generate image aesthetic description texts that combine semantic accuracy and artistic expression through deep learning algorithms. In response to the problem that there is no review article in this field, this article systematically reviews the technical development context in the field of image aesthetic caption, focusing on analyzing the technical route of image aesthetic caption based on traditional deep learning architecture, namely convolutional neural networks and recurrent neural networks, and also introduces the technical application of multimodal models in the field of image aesthetic caption in recent years. This article covers the main technical methods, data sets, evaluation indicators, and future development trends in the field of image aesthetic caption and analyzes the challenges and opportunities of current research. We hope that our review can be a reference for future research in the field of image aesthetic caption.

Accelerators for pose estimation have a wide range of application scenarios at the edge. However, the complexity and variability of the edge environment and the limited power consumption restrict the application and deployment of conventional devices such as GPUs at the edge. In response, this paper proposes a low-power bit-pose estimation accelerator architecture suitable for application in multiple scenarios, and this design supports the mapping of different pose estimation networks to accelerators. In addition, we also develop an accompanying operator library to support new pose estimation networks. In order to cope with the resource- and power-scarce scenarios at the edge, the number of operators in the library can be flexibly configured before accelerator deployment, and the power consumption can be dynamically adjusted by dynamically adjusting the operating frequency of the accelerators in real applications. Finally, we deployed the proposed architecture on Virtex UltraScale+ VU9P and tested it using PVNet as well as MobileNetV2. The experimental results show that the proposed architecture can achieve performance similar to that of current state-of-the-art dedicated accelerators with lower power consumption and greater versatility while ensuring real-time performance and accuracy.

Facial Expression Recognition (FER) plays a significant role in daily life. However, prevalent challenges in expression datasets - including label noise, substantial intra-class variations, and small inter-class differences - severely impact model performance. To address these issues, this paper proposes a dual-branch facial expression recognition network framework. The framework models the similarity distribution of expression data through data modeling and introduces label distribution correction to handle label noise. To mitigate facial identity interference in expression recognition that causes large intra-class variations and small inter-class differences, an average expression anchoring module is employed. Extensive experiments conducted on the RAF-DB dataset demonstrate the model's superior performance, achieving accuracies of 89.09%, 88.62%, and 86.27% on datasets with test noise ratios of 10%, 20%, and 30% respectively, outperforming current state-of-the-art facial expression recognition models in accuracy

Pansharpening is the fusion of panchromatic (PAN) image with multispectral (MS) image to obtain high spatial resolution multispectral (HRMS) image. Due to the limitations of convolution operations and the diversity of remote sensing image features, multi-scale remote sensing pan-sharpening methods cannot effectively establish the connection between features at different scales. In order to establish the connection between different features, we use the "memory" mechanism of GRU and introduce it into the task of Pansharpening of remote sensing image, Establish connections between features of different scales and features at different levels, eliminate unnecessary noise and information redundancy in the process of feature extraction, while retaining important information. Specifically, we proposed a progressive fusion block, in which we proposed a Multi-Scale Memory Interaction Fusion Block and an Adaptive Feature Fusion Block. The former fully extracts features of different scales and establishes the connection between features of different scales, and then The shallow features are fused in a progressive manner to extract features of different depths, while establishing connections between features of different depths. Finally, the Adaptive Feature Fusion Block is used to adaptively fuse the shallow features and deep features to generate a sharpened HRMS. Extensive experiments prove that our proposed method is superior to existing state-of-the-art Pansharpening methods.

In recent years, the detection of small targets in fields such as drone aerial photography and industrial quality inspection has faced challenges such as sparse target pixels, large background noise, and drastic scale changes. Traditional algorithms have problems such as weak feature extraction ability, low multi-scale fusion efficiency, and high computational resource consumption, resulting in insufficient detection accuracy. Specifically, shallow features are prone to losing details, the feature pyramid is insensitive to small targets, and complex networks have difficulty balancing accuracy and real-time performance. To address these issues, this paper proposes an improved algorithm and lightweight solution based on YOLOv7, enhancing detection performance from three aspects: feature enhancement, network optimization, and loss function. This paper designs the CBFR module to dynamically adjust the fusion weights of deep and shallow features, optimizes the detection head structure and adds a small target detection layer to enhance the ability to capture local details; combines the bidirectional routing attention mechanism to focus on key feature regions, and proposes a NWD and IOU composite loss function to balance sample weights. Extensive experiments were conducted on the VisDrone dataset. The experiments show that the improved model maintains the advantages of lightweight while significantly improving detection accuracy and recall rate.

This survey comprehensively examines the role of group dynamics in advancing data analysis and representation learning within Location-Based Social Networks (LBSNs). We systematically review methodologies addressing core challenges such as spatial sparsity, social-location coupling, and computational inefficiency, with a focus on group-aware frameworks that transcend traditional individual-centric models. Key innovations discussed include feature-based group partitioning with transformer networks, social influence integration, adaptive modeling for temporary groups, and dynamic graph representation learning. We critically analyze their effectiveness in enhancing recommendation accuracy, scalability, and interpretability. Applications in urban computing (e.g., crowd flow prediction) and business intelligence (e.g., cross-platform consumer behavior alignment) further illustrate the practical utility of group-aware approaches. The survey identifies unresolved issues—such as real-time adaptability and ethical data governance—while proposing future directions for integrating federated learning, cross-domain alignment, and privacy-preserving mechanisms.

With the continuous development of industrial automation and intelligent technology, the analysis and modeling of industrial production data has become a key means to improve production efficiency and quality. This study explored how to select appropriate models for data processing and optimization in actual production environments by modeling and analyzing the production data of a manufacturing enterprise. Combining relevant domestic and foreign literature, we proposed a comprehensive modeling framework based on machine learning and statistical methods, and designed a highly targeted processing flow, providing a replicable solution for similar enterprises.