Text-to-image generation is an increasingly fast-paced field of generative artificial intelligence, consisting of synthesizing images of high quality and semantic consistency based on natural language descriptions. In this paper, we give an extensive overview of the approach to text-to-image generation using deep learning, including the most common core model families, architecture designs, training approaches, and evaluation systems. We discuss the paradigms of the generative adversarial networks (GANs), variational autoencoders (VAEs), transformer-based designs, and diffusion models, with the last one representing the state of the art in image generation models. The review also discusses key aspects of pipelines such as text encoding, cross-modal alignment, mechanisms of attention, and decoding images. Popular datasets, methods, and metrics of evaluation, including Fréchet Inception Distance (FID) and CLIP-based similarity, are discussed. The application domains that involve creative content creation, medical imaging, education and industrial design are critically discussed. Despite significant advances, various issues still exist, such as low stability in training, excessive computational complexity, amplification of bias, generated images, and text–image alignment errors. Moral and social issues, such as misinformation, intellectual property, and equity, are critically examined. Lastly, we present future research directions to more controllable, more efficient and more interpretable text-to-image systems, focusing on multimodal foundation models and human–AI collaborative design.

Intelligent Reality (IR) systems deployed on resource-constrained edge devices demand efficient and reliable visual perception for real-time applications. This paper benchmarks five compact deep learning architectures —MobileNetV2, EfficientNetB0, MobileNetV3Small, NASNetMobile, and DenseNet121 — on the CIFAR-10 dataset using a unified transfer learning pipeline with frozen ImageNet-pretrained weights and a custom classification head over 50 training epochs. DenseNet121 achieved the highest training accuracy (96.18%), reflecting strong feature representation, while NASNetMobile demonstrated the most stable generalization with a validation accuracy of approximately 77%. MobileNetV3Small offered the lowest computational footprint (~1.01M parameters, 66M FLOPs), making it the most suitable candidate for extreme edge deployments. EfficientNetB0 failed to converge under the frozen-backbone configuration, attributed to a resolution mismatch at 32×32 input rather than an intrinsic architectural deficiency. Evaluation encompassed parameter efficiency, convergence behavior, validation stability, macro-average ROC curves, and per-class confusion matrix diagnostics. To the authors' knowledge, this is the first study comparing these five architectures within a single frozen-backbone pipeline explicitly targeting IR edge deployment, providing an empirical framework to guide architectural decisions for real-time visual perception systems.

Foundation models represent a transformative paradigm in radiological image analysis, offering large-scale, versatile systems that move beyond traditional task-specific AI approaches. While prior surveys have summarized their capabilities in medical imaging, a critical gap remains—the translation from research innovation to sustainable clinical radiology practice. This review directly addresses the implementation gap by presenting subspecialty-specific deployment strategies, real-world performance benchmarks from clinical settings, and practical PACS/RIS integration protocols informed by deployments across 15 healthcare systems. Rather than cataloging model architectures or theoretical advances, we focus on how radiology departments can effectively and successfully implement foundation models in practice, providing actionable, evidence-based guidance for clinical teams responsible for making adoption and deployment decisions. We explore architectures ranging from Vision Transformers to multimodal systems, their pre-training strategies, and adaptation techniques for medical imaging tasks. Clinical applications across multiple imaging modalities have demonstrated significant advances in lesion detection, disease classification, and automated reporting. Despite this progress, challenges remain in achieving robust clinical validation, navigating regulatory approval, and addressing ethical implementation. In this review, we evaluate key performance metrics and examine integration challenges, with particular attention to issues of data heterogeneity, interpretability, and computational efficiency. This review uniquely integrates an imaging informatics perspective, framing foundation models within the full lifecycle of medical imaging data from acquisition to clinical integration. We also highlight emerging solutions and outline future directions, including next-generation architectures, applications in personalized medicine, and strategies to expand accessibility in resource-constrained settings. We further highlight practical solutions for computational efficiency and accessibility, such as edge computing and model compression, vital for broader institutional adoption. This review provides clinical teams—radiologists, administrators, and imaging informatics specialists—with actionable guidance on foundation model implementation, while also serving as a reference for researchers exploring applied AI translation in radiology.

This work aims to implement a multi-feature intrusion detection system for the CAN bus. As vehicle technologies become more advanced, automated, and connected, their electronic systems become increasingly vulnerable to cyberattacks. To address these risks, an effective intrusion detection system is crucial. We propose combining two detection methods: Rule-based Intrusion Detection and Timing ECU Fingerprinting. This integration enhances detection capabilities by compensating for the limitations of each approach individually. Testing was conducted on an embedded board with typical automotive computational power (AURIX TC375Lite) using an experimental prototype to simulate realistic data traffic.

Stroke is a leading cause of disability and mortality worldwide, necessitating early detection for effective intervention. This study introduces a novel, mobile-enabled solution for early stroke detection, leveraging a lightweight deep learning (DL) approach to identify acute and non-acute stroke symptoms from facial features in real time. The proposed system utilizes the YOLOv8n model, a state-of-the-art object detection architecture, which has been fine-tuned on a custom dataset tailored for stroke-related facial anomalies. To ensure compatibility with resource-constrained devices, the trained YOLOv8n model was converted to TensorFlow Lite, a framework optimized for mobile deployment. The system is integrated into an Android mobile application using Flutter, a cross-platform development framework, enabling seamless execution and real-time video streaming from the device's camera. This cutting-edge implementation allows for continuous health monitoring, providing users with immediate feedback on potential stroke symptoms. The lightweight nature of the TensorFlow Lite model ensures efficient performance on mobile devices without compromising accuracy. Experimental results demonstrate the system's ability to detect stroke-related facial asymmetries and anomalies with high precision, making it a promising tool for early diagnosis and timely medical intervention. By combining advanced DL techniques with mobile technology, this work paves the way for accessible, real-time health monitoring solutions, particularly in remote or underserved areas where immediate medical attention is often unavailable.

The early detection of fire and smoke is essential for mitigating human casualties, property damage, and environmental impact. Traditional sensor-based and vision-based detection systems frequently exhibit high false alarm rates, delayed response times, and limited adaptability in complex or dynamic environments. Recent advances in deep learning and computer vision have enabled more accurate, real-time detection through the automated analysis of flame and smoke patterns. This paper presents a comprehensive review of deep learning techniques for fire and smoke detection, with a particular focus on convolutional neural networks (CNNs), object detection frameworks such as YOLO and Faster R-CNN, and spatiotemporal models for video-based analysis. We examine the benefits of these approaches in terms of improved accuracy, robustness, and deployment feasibility on resource-constrained platforms. Furthermore, we discuss current limitations, including the scarcity and diversity of annotated datasets, susceptibility to false alarms, and challenges in generalization across varying scenarios. Finally, we outline promising research directions, including multimodal sensor fusion, lightweight edge AI implementations, and the development of explainable deep learning models. By synthesizing recent advancements and identifying persistent challenges, this review provides a structured foundation for the design of next-generation intelligent fire detection systems.

Driver distraction and inattention are major contributors to traffic accidents, making reliable real-time driver state detection essential for intelligent transportation systems. Although deep learning (DL) has significantly improved detection accuracy, conventional models remain computationally demanding and are often unsuitable for deployment on embedded or low-power hardware. To address these limitations, this study benchmarks seven lightweight architectures—Efficient-Tiny, ESPNetv2-Small, MCUNet, Micro-MobileNet, PhiNets, ShuffleNetLite, and SqueezeNetv11—within a unified evaluation framework tailored for embedded deployment. Experimental results demonstrate that ESPNetv2-Small and SqueezeNetv11 achieved the highest accuracies of 99.50

The Detection Transformer (DETR) has emerged as the dominant paradigm in the field of object detection due to its end-to-end architectural design. Researchers have explored various aspects of DETR, including its structure, pre-training strategies, attention mechanisms, and query embeddings, achiving significant progress. However, high computational costs limit the efficient use of multi-scale feature maps and hinder the full exploitation of complex multi-branch structures. We examine the negative impact of multi-scale features on the computational cost of DETRs and find that introducing long sequence data to the encoder is suboptimal. In this work, we aim to further push the boundaries of DETR's performance and efficiency from the model structure perspective, thus developing the fusion detection Transformer (F-DETR) with heterogeneous scale multi-branch structure. To the best of our knowledge, this is the first explicit attempt to integrate multi-scale features into the end-to-end DETR structure. Specifically, we propose a multi-branch structure to simultaneously utilize feature maps at different levels, facilitating the interaction of local and global features. Additionally, we select certain joint latent variables from the interactive information flow to initialize the object container, a technique commonly used in query-based detectors. Experimental results show that F-DETR achieves a 43.9 % AP using 36 training epochs on the popular public COCO dataset. Furthermore, our approach demonstrates a better trade-off between accuracy and complexity compared to the original DETR.

Nowadays, automatic modulation classification (AMC) plays an essential role in the cognitive radio based non-cooperative wireless communication system. Although numerous deep learning models have been developed for AMC, it remains challenging to effectively recognize various modulation schemes in complicated signal-to-noise ratio (SNR) conditions. In this paper, we introduce a robust AMC method based on the asymmetric trilinear attention net (Tri-Net) with noisy activation function. In Tri-Net, the asymmetric trilinear representation module is developed to deal with various channels of received radio signals and extract rich features to improve the generalization ability. Then a hybrid coding module with squeeze and excitation (SE) blocks-based attention mechanism is designed to help the model adapt to fluctuating SNRs. Finally, the predicted modulation schemes can be output through the classification module. During the training process, the noisy rectified linear unit (ReLU) is proposed guiding the model to explore the convergence position closer to the global optimum. Extensive experiments on practical and simulation communication applications demonstrate that Tri-Net achieves superior classification performance compared with a series of state-of-the-art deep learning models, especially at low SNRs.

This paper examines the challenges encountered in mosques activities management and explores the potential use of internet of things. To this end, qualitative methods, surveys and semi-structured interviews, were conducted. The results reveal key challenges, which are related to mosque accessibility, communication, and engagement in decision-making. As a part of the proposed solution, the Mehrab system incorporates a dedicated social networking, smart locks, and smart key delegation. Mehrab was developed using an Agile methodology that involves potential users throughout the development process to ensure the inclusion of essential features aligned with the system’s objectives. Mehrab was tested by 400 participants who performed a set of tasks that covered the core functionalities and then answered questionnaires to evaluate application objectives achievement, usability, and technology acceptance. Statistical analysis revealed that participants’ evaluations of system objectives differed significantly from a neutral benchmark, indicating a positive perception. A system usability scale survey was used to assess usability, and we achieved an excellent result of 94.23, exceeding the threshold to pass the test. To assess technology acceptance, the technology acceptance model was integrated with institutional theory, offering a robust framework for understanding acceptance factors in religious settings. Regression analysis showed that perceived usefulness, attitude toward using technology, and behavioral intention were the strongest predictors of actual use which commonly interpreted as indicators of technology acceptance, while institutional factors, including normative and coercive pressures, also had significant influence. Moreover, sentiment analysis was employed to measure users’ opinion, which indicated that 94.5% of participants had positive opinions.

With the continuous advancement of wireless communication and the emergence of new communication scenarios, channel estimation, as a core component of wireless system design, has become increasingly significant. This paper reviews important advancements in channel estimation within wireless communication systems, including applications in single-input single-output (SISO), multi-input multi-output (MIMO), orthogonal time frequency space (OTFS), orthogonal frequency division multiplexing (OFDM), and the latest reconfigurable intelligent surface (RIS) systems. We first revisit traditional channel estimation methods, such as least squares (LS), minimum mean square error (MMSE), and compressed sensing (CS), and detail their fundamental principles and scopes of application. Subsequently, we discuss how deep learning techniques offer new perspectives and solutions for channel estimation through models like convolutional neural network (CNN), recurrent neural network (RNN), generative adversarial network (GAN), long short-term memory (LSTM), and graph neural network (GNN), particularly in terms of their potential to handle complicated and dynamic environments. Additionally, we analyze the advantages and disadvantages of these methods in emerging scenarios, including RIS-assisted communications, vehicular networks, indoor positioning, sensing mobile networks, and satellite communications. We also address current methods for evaluating channel estimation performance and highlight the importance of standardization and open data in advancing the field. Finally, we summarize potential future directions for channel estimation and consider its prospects in sixth-generation (6 G) wireless communication systems, aiming to provide a comprehensive technical reference on channel estimation and promote the design of efficient and intelligent wireless communication systems.

The automatic diagnosis of lung cancer via artificial intelligence faces two hotspot issues: (1) insufficient data and (2) excessive redundant information, which make it difficult for convolutional neural networks (CNNs) to learn discriminative information of lung cancer. In this paper, we present the reconstruction error based implicit regularization method (REbIRM) that regularizes CNNs at the loss layer. During each training iteration, the reconstruction errors introduced by the two-stage discriminative auto-encoder are used to sharpen the generalization ability of deep CNNs by improving the decision boundary. In the application process, the trained deep CNN is used for completing computed tomography (CT) diagnostics. The main clinical benefit of our approach is that it is domain independent, requiring no specialized knowledge, and can therefore be applied to different types of datasets. To the best of our knowledge, this is the first attempt to implicitly regularize CNNs based on the reconstruction errors. Finally, experimental results on three CT image classification datasets show that REbIRM can achieve impressive results and that, in conjunction with Dropout, it obtains the state-of-the-art performance. REbIRM is also robust to the selection of hyper-parameters and only has the sublinear influence on the convergence of deep CNNs. Besides, empirical and theoretical evidence are provided to indicate that REbIRM prefers to converges in a constrained parameter space with flatter minima, which explains why it can generalize to new data. Finally, the nature of REbIRM is further explored through visualization techniques to analyze how it works in training deep CNNs.

Drowning is one of the leading causes of unintentional injury-related deaths worldwide, particularly among children. Lifeguards often fail to detect drowning in their early stages, leading to delayed interventions. To address this critical issue, we propose an intelligent video surveillance system leveraging deep learning techniques for early drowning detection. The system is designed to automatically monitor swimming pools, sea and rivers and other aquatic environments in real-time, identifying potential drowning incidents before they escalate. Our approach utilizes real-time object detection models, trained, and validated using a dataset composed of diverse swimming scenarios. We enhance the dataset with various augmentation techniques to improve the model's robustness in different lighting conditions, camera angles, and environmental variations. Key performance metrics, including mean Average Precision (mAP), Recall, Precision, and F1-score, are employed to evaluate the effectiveness of the proposed solution. The model is deployed on edge devices such as NVIDIA Jetson Nano and Jetson Xavier, offering high computational efficiency while maintaining real-time detection capabilities. Through extensive experiments, the system achieved promising performance metrics, with a high mAP and F1-score, confirming its potential for reliable early drowning detection. This research presents a cost-effective, scalable solution aimed at improving the safety of aquatic environments, reducing the risk of drowning incidents, and facilitating timely interventions. By deploying on low-power edge platforms, the proposed system ensures accessibility and efficiency, contributing to enhanced public safety.

This research presents an innovative brain tumor detection and localization approach using the advanced deep learning model, YOLOv9. The superior performance and processing capabilities of this model has been leveraged in this study to address critical challenges in brain tumor detection and localization using medical imaging. The YOLOv9 model has been meticulously trained on a comprehensive dataset of annotated brain MRI scans, achieving remarkable precision in identifying and localizing tumors of various sizes and types. Through extensive experiments, the model has demonstrated marked improvements over previous YOLO versions and other state-of-the-art methods, particularly in detection speed and localization accuracy. The findings suggest that YOLOv9 can substantially enhance diagnostic workflows, and offer a robust tool for early and accurate tumor detection. This advancement holds promise for improving patient outcomes and streamlining medical image processing, potentially setting a new standard in applying deep learning in healthcare.

As an essential technology for spectrum sensing and dynamic spectrum access, automatic modulation classification (AMC) is a critical step in intelligent wireless communication systems, aiming at automatically recognizing the modulation schemes of received signals. In practice, AMC is challenging due to the influence of communication environment and signal parameters, such as unknown channels, noise, symbol rate, signal length, and sampling frequency. In this survey, we investigated a series of typical AMC methods, including key technology, performance comparisons, advantages, challenges, and future key development directions. According to the methodology and processing flow, AMC methods are divided into three categories: likelihood‐based (Lb) methods, feature‐based (Fb) methods, and deep learning methods. The technical details of various types of methods are introduced and discussed, such as likelihood distributions, artificial features, classifiers, and network structures. Then, extensive experimental results of state‐of‐the‐art AMC methods on public or simulated datasets are compared and analyzed. Despite the achievements that have been made, there are still limitations of the individual methods, including generalization capability, reasoning efficiency, model complexity, and robustness. In the end, we summarized the severe challenges faced by AMC and key future research directions.

Artificial intelligence (AI)-based object detection in radiology can assist in clinical diagnosis and treatment planning. This article examines the AI-based object detection models currently used in many imaging modalities, including X-ray Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Ultrasound (US). The key models from the convolutional neural network (CNN) as well as the contemporary transformer and hybrid models are analyzed based on their ability to detect pathological features, such as tumors, lesions, and tissue abnormalities. In addition, this review offers a closer look at the strengths and weaknesses of these models in terms of accuracy, robustness, and speed in real clinical settings. The common issues related to these models, including limited data, annotation quality, and interpretability of AI decisions, are discussed in detail. Moreover, the need for strong applicable models across different populations and imaging modalities are addressed. The importance of privacy and ethics in general data use as well as safety and regulations for healthcare data are emphasized. The future potential of these models lies in their accessibility in low resource settings, usability in shared learning spaces while maintaining privacy, and improvement in diagnostic accuracy through multimodal learning. This review also highlights the importance of interdisciplinary collaboration among artificial intelligence researchers, radiologists, and policymakers. Such cooperation is essential to address current challenges and to fully realize the potential of AI-based object detection in radiology.

There is growing interest in Cyber Threat Intelligence (CTI) as a proactive defense strategy against cyberattacks. Indicators of Compromise (IoCs) are a key component of CTI, enabling organizations to effectively detect, mitigate, and respond to ongoing threats. Although many CTI feeds provide valuable insights into malicious activities by sharing IoCs, they often lack the necessary coverage to support timely and accurate defense against evolving cyberattacks. Moreover, most CTI feeds require organizations to invest in paid services for full access to critical data. This study introduces an open-source framework that leverages a large language model and a novel feature set to extract and rank CTI data based on the credibility of its source, with a particular focus on the X platform. The solution offers features not typically available in existing feeds, such as real-time detection, customization, open-source, and transparency in prioritizing IoCs. To this end, we fine-tune a BERT model to classify whether tweets contain IoC-related content and compare its performance with a suite of traditional machine learning classifiers that utilize various natural language processing techniques. The experimental results indicate that the fine-tuned model with an extended tokenizer and fully trainable layers achieved 99.36% accuracy, outperforming both existing solutions reported in the literature and the evaluated traditional machine learning models. Result represents 14.76% increase in F1-Score compared to the best-performing solution. Moreover, the study examined the collected IoCs from various aspects, uncovering that different credible IoCs were reported on X, some of which were not identified by some other CTI feeds.

A major obstacle to developing accurate deep learning (DL) models for brain tumor classification is the scarcity of labeled magnetic resonance imaging (MRI) data. To address this limitation, this study investigates the use of generative adversarial networks (GANs) for photorealistic brain tumor MRI augmentation, evaluating four popular GAN models, including deep convolutional GAN (DCGAN), conditional GAN (cGAN), lightweight variant of StyleGAN (StyleGAN-lite) and Wasserstein GAN with gradient penalty (WGAN-GP). The models are evaluated using multiple quantitative measurements, including Fréchet inception distance (FID), Inception score (IS), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR), as well as qualitative measurements. The findings demonstrated that WGAN-GP achieved the best overall balance, showing superior performance in FID and IS, especially in later training. StyleGAN-lite produced the highest PSNR, indicating strong pixel-level fidelity but reduced diversity. cGAN generated consistent and structurally reliable images, while DCGAN provided a robust baseline. This research highlights the potential of GAN-generated data, especially WGAN-GP model, to significantly reduce data scarcity, increase dataset diversity, and improve the generalization of downstream diagnostic models for clinical application in neuro-oncology.