Graph neural network (GNN) models are capable of capturing the intrinsic structure and semantic relationships within data and this mechanism grants them substantial potential advantages in the field of computer vision. However, GNN models lack the ability to learn local fine-grained features, which limits their current applications in computer vision. Given the superior capability of convolutional neural networks (CNNs) to learn local detail features, this paper proposes a new module that integrates graph computation layers and convolution computation layers, aiming to learn patch representations that encompass both global and local features, thereby enhancing the performance of GNN models in computer vision tasks. Specifically, (1) the image is divided into patches which are treated as graph nodes, and nodes are connected to their neighbors using KNN, thus representing the image as a graph structure; (2) a module (GraphConv) consisting of a graph network computation layer and a convolution network computation layer is designed, where the former encodes large receptive field information based on the connections between nodes, and the latter learns the fine-grained features within patches using small convolution kernels; (3) the GraphConv module is stacked in a deep model using a residual structure. On the ImageNet-1K dataset and PASCAL VOC object detection dataset, our method significantly improves the performance of the original GNN model. The PyTorch code is available at https://github.com/yuan0038/GraphConvNet.
Although large language models (LLMs) have demonstrated impressive capabilities in generating coherent and fluent text, they often produce irrelevant content when tasked with specialized domains. To address these challenges, Retrieval-Augmented Generation (RAG) combines retrieval and generation processes to enhance the relevance, accuracy, and diversity of LLM responses. However, naive RAG approaches often struggle with precision and recall during the retrieval phase, leading to the selection of misaligned or irrelevant chunks, and sometimes missing critical information. In this work, we propose a novel approach that integrates both word-level and sentence-level retrieval techniques to optimize the retrieval process. By improving the alignment of retrieved information, our method addresses these precision and recall issues more effectively than traditional retrieval approaches. Extensive experiments on benchmark datasets show that our method is significantly superior to existing retrieval strategies, reducing computational overhead and improving accuracy. Our results not only highlight the effectiveness of advanced retrieval strategies in improving LLM performance but also demonstrate their practical implications for scaling NLP systems in real-world applications.
Ensuring safe and accurate drone landings is a critical challenge in the development of autonomous drone systems. This study presents a two-fold approach combining granular semantic segmentation with high-level binary safety classification of aerial imagery. Our method first performs detailed segmentation of environmental elements, then categorizes these segments into "safe to land" and "unsafe to land" zones, providing a more comprehensive and practical solution for autonomous landing decisions. We employed three state-of-the-art deep learning algorithms to perform semantic segmentation on a newly-released dataset of aerial footage from residential neighborhoods. Our models achieved a multiclass Mean Intersection over Union (mIoU) score of 0.82 for detailed segmentation and a binary mIoU score of 0.90 for the critical safe/unsafe landing classification. Among the models tested, U-Net with MobileNetV3 encoder stood out by its better generalization performance, faster training times, and a more compact model size. This study highlights the potential for real-world drone applications, emphasizing its suitability for ensuring safe landings through enhanced environmental mapping and classification.
In recent years, public concern regarding food safety has significantly increased due to the recurrent occurrence of food safety incidents. This paper introduces a comprehensive methodology for assessing food health risks using Fault Tree Analysis (FTA). Initially, FTA is employed to quantify the probability of contamination, thus determining the likelihood of risk occurrence. Subsequently, a diverse set of risk indicators is integrated, and a weighted assessment—incorporating both subjective and objective weighting schemes—is conducted to derive a comprehensive measure of risk severity. The risk matrix approach is then utilized to classify levels of risk likelihood and severity, effectively synthesizing contamination probability with environmental health risk to determine an overall risk value. The proposed framework is applied to evaluate cadmium contamination in rice samples collected in 2019 from the provinces of Hunan, Hubei, and Jiangxi. The findings indicate that Jiangxi represents a high-risk area, Hunan a relatively high-risk area, and Hubei a comparatively lower-risk area.
In this paper, we introduce and experimentally demonstrate a fault detection method for the air conditioner in the EMU train, which usually generates abnormal sounds when the faults occur. The technique utilizes the Support Vector Machine (SVM) to recognize the time-frequency characteristics of the sound generated by the air conditioner, which is extracted using the short-time Fourier transform (STFT). In addition, a detection system is built for the experiment, in which an FPGA and an embedded system are employed for sound acquisition and acoustic signal process, respectively. The penalty factor of the SVM model is optimized and accuracy of 0.977 is achieved. Our method will improve the efficiency and accuracy of fault detection for the air conditioner in the EMU train.
To enhance the precision of the robotic arm dynamics model and ensure the consistency of unidentified parameters across various operating conditions, a parameter identification method based on the multi-objective particle swarm algorithm is proposed. Using the LuGre friction model, the dynamics model of the robotic arm are established. Parameters are identified by minimizing the difference between simulated and actual displacements under 30°and 50°conditions. Verification is conducted at 40°, demonstrating reduced mean square errors of the angular displacement curves by 9.31% and 9.10% for the initial two conditions. These results confirm the effectiveness of the parameter identification method in improving the model's accuracy.
Research focus on the trade-off between diagnostic accuracy for Alzheimer's disease (AD) and time management for diagnosis is very limited. This study proposes a novel two-stage feature selection framework integrating variance inflation factor (VIF) and NSGA-II optimization algorithm to obtain a set of sub-items of cognitive and neuropsychological assessments (CNAs) with high diagnostic accuracy and short managing time. What sets the proposed framework apart is that it 1) mitigates multicollinearity issues caused by multiple types of feature selection algorithms employed on multiple CNAs and 2) has the multi-objective optimization capacity of searching Pareto solutions in a reduced decision space. Crucially, we design a VIF-based fast forward search for feature selection which takes into account the cost of managing time for CNAs. Experimental results demonstrate a significantly shortened managing time of 905 seconds (reduced by two-third), with a slightly improved AUC performance of 0.9113 for classifying cognitively normal controls, mild cognitive impairments, and AD patients based on a set of selected combination of sub-items compared to their involved full CNAs, thereby offering promising prospects for accurate diagnosis and fast assessment in clinical practice.
Pain is a complex phenomenon that incorporates both physical sensations and emotional responses. The use of automated pain assessment is crucial in order to develop effective medical diagnostic systems for pain management. Thus, it is important to conduct a study on the outcomes achieved by utilising deep learning algorithms for the detection of pain expression. This study aims to provide reliable and unbiased methods for the automated assessment of pain. The aim of this systematic review is to discuss the models, methods and data types used to build the foundation of deep learning-based automated pain assessment systems, with a focus on analysing improved strategies and methods based on deep modelling techniques used to enhance feature extraction accuracy and the accuracy of pain level assessment. As a result, the literature explores the application of facial expression recognition techniques in the field environment for addressing various challenges in clinical testing. Consequently, the systematic review identifies the limitations of the current research on automated pain assessment and offers an outlook on the potential future research directions.
The QA system can provide various professionals with fast and accurate knowledge support, significantly improving work efficiency. The introduction of LLMs has further enhanced the accuracy and efficiency of QA systems. A QA system based on LLMs for power transformers can effectively improve the efficiency of fault diagnosis and the resolution of technical issues in the power transformer domain. This study first proposes an automatic dataset construction method for QA based on LLMs, through which a QA dataset of basic knowledge about power transformers is obtained. This dataset then performs low-rank adaptive fine-tuning on the LLM. Subsequently, an external knowledge base covering fundamental knowledge and fault cases of power transformers is built. Finally, with the aid of prompt texts, knowledge responses were generated by RAG in combination with the fine-tuned LLM. Experimental results show that compared with the QA systems driven by general LLMs, the method proposed in this paper generates more concise and professional responses, thus promoting, to some extent, the application of LLMs in the power domain.
This paper addresses the significant challenges posed by road safety due to rapid urbanization and increasing vehicular traffic. High-definition (HD) semantic maps are essential for improving decision-making and safety in autonomous vehicles and intelligent driver assistance systems, which require detailed road environment information. Current multi-camera perception systems face limitations in efficiently integrating data, resulting in inconsistent spatial understanding. To tackle these issues, this propose a novel multi-scale feature optimization mechanism EPPAMap, including the efficient depthwise separable feature network(EDSFN) to improve semantic understanding. Additionally, we introduce the parallel attention aggregation module (PAAM) to capture richer contextual information and enhance feature aggregation. Experimental results demonstrate superior adaptability and performance in complex road scenarios, providing a solid foundation for HD semantic map construction and real-time traffic perception.
The use of multimodal sentiment analysis is now gaining its popularity to analyze a user's emotions and feelings in a social media platform. However, the correlations between visual and textual content have been neglected leading to disparity of results. Recently, many of the articles in relation to the field of study concentrate on the analysis of unimodal concept and when there is the study of multimodal analysis it raises the issues of error due to the concern of heterogeneous description between text and image data. Motivated by this status quo, this paper aims to propose a novel improved deep multi-view attentive network in image and text data. The study applies three core processed for: data feature extraction; training, validating, and testing data; and the interpretation of the classified multimodal sentiments. The novelty of the approach is shown in the first and last phase of the study. The initial phase of the study does not only produce a classified sentiment but it also produced sub-classifications to remove heterogeneous information for image and text data. Secondly, image data includes subsets of image sizes to optimize results from different layers and regions. In this, the study achieved 93% of accuracy and though it's a bit lower from the other studies related into, it shows less concern with overfitting of results that is based on the heterogeneous description between image and text data. The process and its analysis demonstrate a superior performance and could be used as a current state-of-the-art technique to evaluate fused multimodal sentiments in social media.
Object counting aims to estimate the number and distribution of people using images or videos. Single-view crowd counting commonly encounters issues such as object occlusion and poor visibility. Existing multi-view crowd counting methods rely heavily on the internal and external parameters of cameras, projecting information into a shared coordinate system, resulting in high computational complexity and limited applicability. Based on these observations, this paper proposes a method where objects are matched to identify new individuals (those heavily occluded from the main view but detectable from auxiliary views), enabling corrections to the count to enhance counting accuracy. Furthermore, we have compiled a multi-view indoor people counting dataset. Ablation experiments and comparison studies conducted on this dataset and the publicly available PEST2009 dataset validate the effectiveness of the proposed approach.
This paper introduces a novel approach to enhancing the performance of Log-Structured Merge (LSM) trees through the implementation of deferred updates. Traditional LSM trees suffer from performance inefficiencies due to the need for hidden reads during replace and delete operations. Our approach eliminates these hidden reads by deferring old data lookups and deletions until the compaction phase. I present a detailed examination of the deferred update algorithm, including its integration with various sublevel compaction strategies and its impact on read and write performance. Theoretical analysis and empirical evaluation demonstrate that deferred updates significantly accelerate operations, achieving up to a tenfold increase in speed compared to conventional methods, particularly under write-intensive workloads. I also explore the implications of this optimization on secondary index management and overall system throughput. Our results suggest that the proposed method not only improves efficiency but also offers a scalable solution for managing large-scale databases with multiple indexes.
Strategic emerging industries play a vital role in overall long-term economic and social development. However, strategic emerging industries face challenges such as insufficient independent innovation capabilities, a lack of strong support in terms of talent and funding, and incomplete institutional mechanisms, resulting in high risks for industries. Therefore, the issue of risk assessment and prevention for emerging sectors of strategic importance has gained attention in the academic community. However, most existed studies of the industry risk mainly focused on summarizing and interpreting risk phenomena, deeply systematic classification and quantitative analysis of various risks were largely unclear. Big data thinking guides industrial innovation by including diverse data. The development of big data thinking can make up for the shortcomings of traditional risk identification methods, such as incomplete data, excessive attention to accuracy, low data utilization, etc. In this study, the industrial risk assessment system was firstly constructed based on the guidance of big data thinking from the perspective of both internal risk and external risk. Then, the entropy weight method and cloud model were combined to evaluate the industry risk. The new energy automobile industry is representative of the development direction of strategic emerging industries. The associated risk was evaluated based on entropy weight-cloud model in China. The results showed that the impact of external risk on the industry was greater than the internal risk, and the industry risk was at a higher medium level. Finally, some suggestions for preventing industry risk were proposed on the basis of risk analysis, such as improving the industry’s own management capacity, promoting product and technological innovation, reducing market risk, etc.
To maintain the functionality of complex networks, recent studies focus on optimizing structural stability through anchor operations. The anchored coreness problem, for instance, seeks to maximize the overall coreness of all nodes by anchoring a selected subset. However, existing approaches overlook the cost associated with node anchoring, a critical factor in budget-constrained scenarios. To address this, we introduce the problem of maximizing coreness gain via incentive allocation, which aims to determine the optimal allocation of incentives to nodes under a given budget and node-specific cost functions, thereby maximizing the network’s overall coreness. We propose a baseline algorithm, GreedyIA, and an efficient heuristic algorithm, Batch-GreedyIA, designed to identify key nodes and their optimal incentive allocations. Batch-GreedyIA improves the computational efficiency of GreedyIA by incorporating a batch processing strategy and a reuse technique. Extensive experiments on eight real-world datasets validate the effectiveness of our model in enhancing network stability under budget constraints and demonstrate the efficiency of the proposed algorithms.
In order to solve the problems of large volume, multiple dimensions and difficult analysis in the big data of geographical indication products (litchi, banana, etc.) of Guangdong province, a big data processing and visualization method based on GIS for geographical indication products of Guangdong province was proposed. By constructing GIS technology to pre-process the data of geographical indication products (litchi, banana, etc.) in Guangdong province, the number of data samples and dimensions are reduced; and then through spatio-temporal grid computing and various visualization techniques, the market development situation of geographical indication products (litchi, banana, etc.) is visually displayed. This proposal can achieve fast, interactive and visual big data analysis and mining of litchi, banana and other agricultural products brands, providing technical support for the analysis of geographical indication agricultural products brands in Guangdong province.
Diversified firms improve performance and competitiveness in the market by efficiently allocating resources among their various products or services, leading to economies of scope. In this study, we focus on product diversification in information technology industry. In detail, we take a sample of text data from 3,623 10-K filings of publicly traded companies in U.S. IT industry from 2000 to 2021, and investigate the development trend of product diversification by using text mining technique. Six key topics for IT firms’ products are identified: "Semiconductor and Electronic Product Manufacturing", "Telecommunications and Communication", "Television and Broadcasting Media", "Software and Internet Services", "IT based Healthcare and Life Sciences", and "Cloud Computing and Cybersecurity". These topics represent the main focus of IT companies in terms of product development and services offered to customers. IT product diversification demonstrates a "U" shaped trend between 2000 and 2021, as indicated by the Hirschman Herfindahl index and Concentric index. This pattern suggests fluctuations in IT product diversification over the specified period, with a notable turning point. This finding highlights the dynamic nature of diversification strategies in the IT Industry.
With the rapid development of smart highways, effective visualization and intelligent analysis of data information are important means to improve management efficiency and gain insight into traffic trends. This paper proposes a highway data information visualization tool based on D3.js and GeoJSON, aiming to achieve intuitive display and dynamic interaction of massive traffic data. The tool designs visualization functions of highway networks, toll stations, service areas and vehicle trajectories through a variety of methods such as basic data of highway networks and dynamic data of vehicle trajectories, and supports real-time monitoring and data analysis. Studies have shown that the tool performs well in scenarios such as road network section display, vehicle trajectory restoration, and abnormal event monitoring, which helps traffic managers quickly understand complex data and thus optimize decision-making strategies. Experimental results verify the effectiveness of the tool and its application potential in highway management.
In recent years, electroencephalogram (EEG) signals have been increasingly used in the automatic identification of brain diseases. Most of the current EEG-based methods for epilepsy classification focused on single dimensional feature extraction techniques in the time domain or frequency domain, respectively. In this paper,we proposed a joint time-frequency domain feature extraction method using tunable-Q wavelet transform and Hjorth parameters. The tunable-Q wavelet transform is used to decompose the EEG signal into different subbands, the optimal subbands are calculated according to the maximum energy criterion and divided into different frequency ranges, and the Hjorth parameters (activity, mobility, and complexity) are calculated on the optimal self-bands of different frequency ranges, which are applied to the machine learning algorithm to realize the automatic epilepsy detection. The experimental results show that the method has a high classification accuracy, determines the frequency range in which seizures are active and further improves the accuracy of automated seizure detection.
According to the development demand of intelligent storage industry, this paper designs an intelligent storage goods data acquisition system based on OneNET. The system collects data of storage goods through the magnetic coupling between readers and tags, and uses sensors to obtain storage environment data. Based on MQTT protocol, the storage data is uploaded to the OneNET cloud platform and published by topic. Users who subscribe to the topic can access the data remotely. The system improves data security, realizes asynchronous operation between data producers and data consumers, and improves system efficiency. A prototype is made and the functional verification of the system is completed, which has a good engineering application prospect.
