All articles published in this issue have undergone a thorough peer review process, and stringent checks for repetition rates have been implemented to ensure the integrity of the content.
Total number of articles in this issue: 14 Total number of pages in this issue: 122
For inquiries regarding the content of specific articles, please feel free to contact the respective authors via their provided email addresses. For questions related to the journal itself, please reach out directly to SUAS Press.
Year
2024
Volume
1
Number
4
Status
Archived
Published
2024 November 2
Articles
Open Access Research ArticlePublished 2 November 2024
Multimodal interaction technologies enhance the planning of human-to-human virtual meetings that involve Call Centers located in one part of the world and customers in other activities through the use of interacting with people regardless of their languages thus addressing problems of culture, equity and more interaction and participation from humans is sought after. Hence, language translation and transcription in real-time is an important aspect of multimodal technologies. In addition to these functions, emotion analysis, natural language processing and sentiment analysis, in their turn, allow the host to understand how participants feel and the main ideas expressed in the audio or video minutes. Also, participants have the opportunity to record their notes, which will in turn be generated in the form of minutes using natural language processing. Thus, turning audio and videos into texts significantly improves the efficiency and quality of meetings with natural human engagement.
Disciplines:Artificial IntelligenceSubjects:Natural Language ProcessingReferences:34
Multimodal generative models have become essential in the deep learning renaissance, as they provide unparalleled flexibility over a diverse context of applications within Natural Language Processing (NLP) and Computer Vision (CV). In this paper, we systematically review the basic concepts and technical improvements in multimodal generative models by discussing their applications across different modalities such as text, images, audio,and video. These models though augment the strength of AI to comprehend and perform complicated tasks by coalescing data from various modalities. In this paper, we investigate how these principles apply to many of the existing mainstream models (including CLIP, DALL·E, Flamingo), and consider their applications in VQA,text-to-image-synthesis; medical image analysis; edutainment content creation & user research developments. This paper also examines the existing difficulties of such technologies including paucity in data availability, modality fusion effectiveness and constraints on computational resources while suggesting pathways for future research. The paper goes on to state privacy parallels between multi-modal generative models (GGMs) calls for a model of safety over responsibility when it comes to technological innovation.
With the rapid development of smart manufacturing technologies, automated production has become an important trend in the transformation of the industrial chain. Among various automation applications, robotic arm grasping and visual inspection systems are the most widely used. This paper focuses on unstructured stacking scenarios and workpiece defect detection, and designs two deep learning-based vision systems. In terms of theoretical research, the study focuses on the fundamental knowledge and technical methods related to robotic arm grasping in unstructured environments and workpiece defect detection. To address the issue of grasping randomly stacked objects, a 2D/3D vision-based robotic arm grasping solution is proposed. This solution employs an eye-in-hand configuration, where RGB and depth images are captured by a stereo camera, and a depth feature extraction branch is added to the Mask R-CNN network to improve the accuracy of object detection and segmentation in complex scenes. For object localization, the segmented results are mapped to a 3D point cloud through RGB-D data registration, and the RANSAC and PCA algorithms are used to extract the target plane and bounding box, thereby obtaining the 6D pose information of the target. Combined with the hand-eye calibration results, the robotic arm can accurately grasp the target. Additionally, taking an automotive one-way clutch as an example, an automated defect detection system based on deep learning is designed. Using an industrial camera to capture images, the system utilizes a semantic segmentation network and a defect classification network to detect the number of teeth, copper sleeve, semicircular piece, and chamfer of the one-way clutch, thereby achieving automatic recognition of part defects. This paper integrates 2D image and 3D point cloud information, combined with deep learning methods, to explore robotic arm grasping and workpiece detection, providing new ideas and solutions for the development of smart manufacturing.
With the continuous development of industrial production, efficient and accurate inspection of mechanical equipment has become crucial for production safety, efficiency, and economic benefits. By collecting and analyzing imaging data from electromechanical equipment, effective online monitoring and fault diagnosis can be achieved, enhancing operational efficiency and accuracy while reducing manual intervention. This is a significant research direction in the field of industrial automation control. This paper begins with the fundamental principles of electromechanical equipment testing, conducting an in-depth study of its working mechanisms and providing a detailed discussion on its design and development. The main content includes the system architecture design, functional module design, and key algorithm design, laying a solid foundation for the research on automated testing of electromechanical equipment.
This paper reviews the current state of research on generative AI both domestically and internationally, exploring its potential applications and challenges across various fields. In education, generative AI enhances students' academic writing skills and learning outcomes by providing personalized learning support. In design, it facilitates personalized and innovative creations, enabling designers to generate novel ideas through algorithms. Additionally, the application of generative AI in psychology reveals the complex relationship between emotion analysis and social behavior, while in computer vision, it advances facial recognition technology. However, with the widespread use of generative AI, ethical and social responsibility issues are increasingly prominent. This paper emphasizes the importance of establishing appropriate regulations and legal frameworks to ensure the authenticity and morality of generated content, making this a key focus for future research. Overall, generative AI is profoundly transforming research and practice in various fields, and future studies must pay greater attention to its social impact and technological responsibilities.
In the process of introducing gesture recognition, it is essential to explore its technical background and implementation methods. Gesture recognition algorithms based on deep learning perform exceptionally well when processing real-time video streams. These algorithms can extract gesture features and classify them to identify user intentions. For instance, analyzing gesture images using Convolutional Neural Networks (CNN) can effectively enhance recognition accuracy and real-time performance. Additionally, combining optical flow methods with object detection techniques allows for real-time tracking of user hand movements, leading to more precise recognition results. Factors such as changes in ambient lighting, cluttered backgrounds, and the diversity of user gestures can all impact recognition accuracy. Therefore, researchers need to continuously optimize algorithms to improve the robustness and adaptability of the system. At the same time, when designing virtual conference assistants, the user interface's friendliness and usability should also be considered, enabling users of varying technical skill levels to use the system with ease.
This paper presents the novel work of developing an intelligent manufacturing framework based on multimodal sensor integration and computer vision. In this paper, we propose a hybrid fusion method that includes both early and late fusion with attention mechanisms to select the most important sensor data. Our system gathers visual, thermal, acoustic, and vibration data and offers accurate and interpretable predictions for fault identification, process enhancement, and product quality. (Liu et al. 2024) We meet the challenge of the opacity of AI systems by using explainable AI methods to help the user comprehend the results of the model. It shows that the proposed system is accurate, efficient, scalable, and can be applied to various types of data. Examples from the industry present real-life experiences and issues that may be encountered when implementing our system in different manufacturing contexts. It presents a new paradigm shift in smart manufacturing systems through the enhancement of efficiency, reliability, and interpretability for future research and industrial development.
Temperature serves as a fundamental indicator of thermal conditions, influencing various natural processes and human activities. This study investigates the relationship between temperature and other meteorological factors, including humidity, wind speed, visibility, pressure, and apparent temperature, using historical weather data from Szeged, Hungary (2006-2016). Employing multiple regression models and advanced machine learning algorithms such as XGBoost and Artificial Neural Networks (ANNs), the research aims to elucidate the linear and non-linear dependencies of temperature on these factors. The findings indicate a significant linear correlation, with XGBoost outperforming traditional regression approaches in predicting temperature variations. This study contributes to enhancing temperature forecasting accuracy, which is crucial for improving quality of life and informing climate-related decision-making processes.
Game theory is a branch of complex system analysis that deals with the mathematical modeling of strategic interactions among rational agents. It has been widely used across various fields, such as economics, social sciences, computer science, and operational research. This paper provides an overview of game theory, discussing its foundational concepts, including non-cooperative and cooperative games, Nash equilibrium, and mixed strategies. The study explores extensive-form, repeated, Bayesian, and coalitional games, highlighting applications in operational research and the implications for decision-making under competition.
Video conferencing has become a crucial tool for global remote collaboration, but existing platforms have limitations in capturing and conveying participants' emotions. Real-time emotion recognition technology, by combining computer vision, deep learning, and temporal analysis, can automatically analyze and identify emotional changes in participants, addressing this gap. This paper first introduces the core technical processes of emotion recognition, including face detection, feature extraction, and emotion classification, with a focus on the technical details of using Convolutional Neural Networks (CNNs) for feature extraction and classification algorithms.[1] To enhance the system's temporal dynamics, the paper also presents methods for capturing emotional changes based on Long Short-Term Memory networks (LSTMs) or Temporal Convolutional Networks (TCNs). In particular, we use a Softmax classifier for probability estimation of emotions, coupled with temporal analysis methods for real-time engagement assessment in meetings. Furthermore, this paper discusses the system implementation under an edge computing and cloud collaboration framework to optimize real-time performance and computational efficiency while proposing security strategies focused on privacy protection, such as homomorphic encryption and federated learning. Through specific application examples, this paper demonstrates the significant role of real-time emotion recognition technology in video conferencing and its potential impact on future remote communication modes.
This paper presents a novel hybrid approach for enhancing personalized search by integrating deep learning techniques with cloud computing infrastructure. The proposed system uses a multi-layer adaptive model augmented with a hierarchical monitoring network to capture user preferences and query semantics. Cloud-based architecture, used for Amazon Web Services, provides the necessary scalability and computing resources for the processing of large-scale research data. The system employs a custom middleware layer for efficient integration of the deep learning component with the distributed cloud infrastructure. An analysis of data on 100 million searches showed significant improvements in search accuracy and user satisfaction. The combined method achieves a 15% increase in Average Precision and a 12% improvement in Cost-effectiveness compared to the state-of-the-art baseline. Scalability analysis reveals the performance, maintaining sub-200ms latency for 95 percent. The system transforms the resource allocation efficiently into a non-volatile operation, demonstrating its potential for real-world deployment. This research contributes to the evolving field of AI-driven search optimization, solving problems in personal accuracy, scalability, and efficiency. The findings have implications for the design and implementation of ongoing research, providing insight into the integration of advanced machine learning with cloud resources.
This paper discusses the relationship between financial fraud risk and credit risk in China's financial market, and points out that financial statement fraud has a serious impact on the fairness and transparency of the capital market, and damages the legitimate rights and interests of investors. Research shows that financial fraud risk may be an important signal of credit risk outbreak, and may further spread through the credit network, resulting in a larger scale of credit risk. This paper reviews the evolution of credit risk assessment models, from early expert analysis methods to modern statistical and machine learning-based methods, including Z-score models, SVM models, and random forests. Special attention is paid to the application of the C5.0 algorithm in credit risk assessment, highlighting its advantages in terms of data characteristics and prediction accuracy. Finally, the ROC curve and KS curve are used to evaluate the prediction effect of the model, which shows that the model has good prediction ability and practical value, and provides a new methodology for financial fraud risk assessment.
Disciplines:Computer ScienceSubjects:Machine Learning in FinanceReferences:35
In this paper, we provide a systematic review of the emerging field on applications for Large Language Models (LLMs) in multimodal learning, especially how such methodologies help improve orchestrated task performance by integrating different modalities like images, text, and audio. Multimodal learning is a field where we combine various types of data to make models learn multiple attributes and generate meaningful outputs. It is widely applied in image captioning, cross-modal retrieval, sentiment analysis, and speech recognition. It reviews the main multimodal learning approaches, such as feature extraction, modality alignment, and fusion strategies (early fusion, late fusion, and hybridization), and the performance of LLMs in cross-modal tasks. It highlights the present technological challenges, emphasizing concerns regarding computational resource utilization, model complexity, as well as a lack of multimodal fusion. Lastly, the article provides some suggestions for future applications on how to better integrate modalities and few-shot learning in cross-modal generation models. It also discusses ways to make multimodal machine translation systems run faster using less distributed computational power.
Disciplines:Computer SciencesSubjects:Large Language ModelsReferences:25
Machine Learning, a pivotal technology within the realm of artificial intelligence, has experienced remarkable progress in recent times. This research offers a thorough and structured presentation of machine learning. It begins with a comprehensive look at the evolution of machine learning throughout history, then zeroes in on dissecting the foundational algorithms that underpin the field. Following this, the study sheds light on the cutting-edge developments in machine learning, with the goal of thoroughly examining its applications across different sectors and contemplating the prospective trajectories for its future.
