Neural Style Transfer (NST) is a deep learning technique to fuse artistic features of image content with another, creating a new synthesized image. Such artistic style (colors, textures, brushstrokes) of the content image is based on Convolutional Neural Networks (CNNs) to achieve real-time performance. But it has limitations with fine-grained user control, slow processing speed for the iterative method, and the inability to maintain temporal quality consistency. Existing diffusion models have evolved to provide real-time/near-real-time performance for image generation and style transfer (Image-to-image conversion). GAN-based models traditionally had issues with flexible style transfer in the case of fixed mapping between two specific domains (training phase). Such stylization improves the intensity interpolation parameter λ used for image-to-image style transfer. This paper proposes "StyleFlex", a GAN-based real-time and controllable architecture that combines a generator with Adaptive Instance Normalization (AdaIN) and a new style-intensity interpolation parameter λ . It also provides user-controllable interpolation and better inference time. Unlike classical CNN approaches, the proposed model employs adversarial training to enhance realism and supports smooth stylization control ranging from content preservation ( λ =0) to full stylization ( λ =1). Experimental results demonstrate improved performance over state-of-the-art approaches in terms of stylistic fidelity and computational efficiency. In the future, such a model can be extended to support multi-style combination, style-specific adaptation, and 3D content adaptation.
Time series forecasting employs historical data to predict future events, finding extensive applications in sectors such as energy, transportation, and climate. Prior models for long-term time series forecasting encountered challenges, including insufficient feature extraction from time series data and an inadequate capacity to capture recent local trends. To address these challenges, we propose a novel model, termed Second-order Season-trend GRU (SndSTGRU), which incorporates a second-order hidden state into the existing GRU framework to enhance its representational capacity. Additionally, it employs a separation method to extract seasonal and trend features from time series data, thereby improving the model's feature extraction capabilities. The SndSTGRU model has been empirically evaluated on several real-world datasets, including ETT, Electricity, Traffic, and Weather. Across the 28 forecasting tasks, SndSTGRU obtains the lowest mean squared error (MSE) in 21 tasks and secures a position among the top two performers in 26 tasks. These empirical results demonstrate that the incorporation of seasonal-trend decomposition and a second-order GRU structure significantly improves the accuracy of long-term time series forecasting.
Alzheimer’s disease (AD), the most common type of dementia, is characterized by a progressive decline in memory and cognitive function that dramatically reduces quality of life in patients. Accurate and early AD detection is essential to timely enable the treatment and delay the progression of disease. To address these challenges, a novel deep learning framework is proposed, termed AD-CNN, specifically optimized for Alzheimer’s disease classification. The proposed framework provides an optimized EfficientNet-B0 where replaces classical CNN with lightweight deep pre-trained network on the Image Net, fine-tuned to adapt the structural and spectral characteristics of MRI data. The model consists of the seven stages of the MBConv with Squeeze-and-Excitation (SE) attention and Swish activation functions that enable recalibration channel efficiently with representation of the non-linear learning. In the convolutional backbone, the network integrates projection of three layer fully connected head with (1280 → 256 → 128 → 5) to batch normalization and dropout for generalization. The framework is trained on optimizer of AdamW with 24 mini-batch and 2 × 10⁻4 learning rate, 30 epochs to ensure the computational efficiency and maintain the high precision of detection. After performing the convergence, 128 dimensional features to final projection layer are extracted and refined further by atom search optimization (ASO) to identify discriminative subset of features. Finally optimized features are used for classification based on the ensemble classifiers. This optimization step effectively eliminates redundant information and keeping significant features to increase the classification robustness. These features fed to Support Vector Machine (SVM), tree based and neural network classifiers to confirm the performance is consistent in different types of classifiers. The experiments are carried out on Alzheimer’s Disease Neuroimaging Initiative (ADNI)-1, ADNI-2, and OASIS-2 datasets to depicts that proposed CNN based framework provides 99.9 https://doi.org/10.5281/zenodo.19248909 ”.
Stock market forecasting plays a significant role due to the highly volatile, nonlinear, and interdependent nature of financial time series. Existing deep learning and statistical models face challenges in effectively capturing both spatial and temporal correlations among stocks. To address these limitations, this paper proposes a novel hybrid forecasting framework. In the proposed strategy, dynamic trading-day data are utilized to construct similarity-based multi-neighbor graphs using price, volume, and volatility indicators. These graphs are processed through message passing to extract spatial characteristics, which are subsequently modeled using a Chaotic Long Short-Term Memory (LSTM) network. The LSTM weights are initialized using the Chen chaotic system to enhance sensitivity to market variability and nonlinear fluctuations. The proposed framework is rigorously evaluated on three financial datasets: NASDAQ, Nifty 50, and the Chinese A-share market. Experimental results demonstrate superior performance compared to several state-of-the-art baseline models, achieving prediction accuracies of 95.82 R^2 ) of 0.971 on the Nifty 50 dataset. These findings highlight the effectiveness of integrating multi-scale spatial reasoning with chaos-based temporal modeling for stock price forecasting.
Cryptocurrency markets represent highly complex and adaptive systems with non-linear and multi-scale dynamics, making them ideal for analysis through the lens of network science and statistical mechanics. This study introduces a robust network-oriented framework for forecasting cryptocurrency price dynamics, leveraging visibility graph-based reconstructions to capture temporal structures and dependencies in market data. Key network features, including centrality measures, community structure, local efficiency, and eccentricity, are extracted to characterize the underlying system’s topology. These network metrics are incorporated into interpretable ensemble learning models, specifically Random Forest and LightGBM, to yield high predictive accuracy while elucidating the critical role of network topology in shaping cryptocurrency price evolution. To ensure interpretability in financial decision-making, this study uses Local Interpretable Model-agnostic Explanations (LIME) to reveal how key network metrics, such as community-based metrics and local efficiency, drive price predictions, providing transparency and reliability in high-risk markets. By bridging network science and econophysics, this work provides a novel and transparent approach to modeling cryptocurrency markets, offering unique perspectives into forecasting challenges, feature significance, and the interplay between structural patterns and market behavior. The findings contribute to advancing the understanding of cryptocurrencies as complex systems within the global financial ecosystem, reinforcing their transformative role in economic dynamics.
The inclusion of electric vehicle (EV) charging station (EVCS) into radial distribution network (RDN) along with renewable distributed generation (RDG) and distributed static compensators (DSTATCOM) allocation by a novel method, namely, quasi oppositional arithmetic optimization algorithm (QOAOA) is the main focus of this study. The inclusion of EVCS increases the load demand and the power loss. Also, the cost of energy loss of the system is increased. For the compensation of the power loss and energy loss cost, RDG, and DSTATCOM are allocated. The RDG is based on two sources, the photovoltaic (PV) module and the wind turbine (WT) which provide environmental benefits. To validate the usefulness of the presented QOAOA algorithm, it is examined on 33-bus, 69-bus, and Portuguese 94-bus systems with four separate cases namely, (i) EVCS placement, (ii) EVCS and DSTATCOM placement, (iii) EVCS and RDG (PV, WT based DG) placement, and (iv) EVCS, RDG (PV, WT based DG), and DSTATCOM placement. To upgrade the overall performance of the RDN, real power loss minimization and annual economic loss reduction are taken as the two main objective functions for this study. The percentage improvement in active power loss for case 2, case 3, and case 4 are 30.94
The Sand Cat Swarm Optimization Algorithm (SCSO) is a new meta-heuristic optimization algorithm inspired by the search and attack behavior of sand cats in nature. To address the shortcomings of the standard sand cat swarm optimization algorithm, such as its tendency to fall into local optima and slow convergence during the optimization process, and to enhance its performance in solving complex problems, this article proposes an adaptive selection dual-population sand cat swarm algorithm with modification strategies(MDPSCSO). First, the HaltonRR2 initialization method is employed in the initial phase to generate a higher-quality and more uniform initial population. Modification strategies are subsequently applied to adjust the best and worst solutions within the population under different fitness landscape conditions, enabling the algorithm to avoid local optima and increase solution diversity during the optimization process. Finally, a dual-population collaboration mechanism is introduced, along with two selection methods, to guide the population toward the correct direction of exploration and accelerate the convergence speed of the population. The combination of these strategies is used to enhance the algorithm’s optimization performance in solving complex problems, and these modifications have been successfully applied to UAV path planning scenarios in complex environments. To validate the proposed algorithm, tests were conducted on the CEC2017 benchmark suite, and simulation experiments were performed in eight 3D UAV path planning scenarios with varying complexities to evaluate the performance of MDPSCSO. The experimental results demonstrate that MDPSCSO outperforms other advanced algorithms in terms of search efficiency, solution accuracy, and convergence speed in solving global optimization problems and that it can effectively generate higher-quality paths for UAVs in complex 3D environments, showing significant potential for practical applications.
Accurate segmentation is crucial for the efficient diagnosis and treatment planning of brain tumors. Automating this process offers substantial advantages for clinicians by reducing the time and potential errors inherent in manual segmentation. This paper introduces RLU-Net-GD, a deep learning-based encoder-decoder architecture designed to address challenges such as image artifacts, poor contrast, class imbalance, and high intra-class variability. The model utilizes ResNet-50 as its backbone, incorporating a residual learning mechanism to mitigate vanishing gradient issues and enhance feature propagation in deep networks. Skip connections between the encoder and decoder pathways ensure the retention of spatial details by combining fine and coarse-grained features. A guided decoder supervises the learning process to enhance the generation of detailed feature maps, while a weighted guided loss function addresses class imbalance. Transfer learning with pre-trained weights also enables effective training with limited data, further boosting segmentation accuracy. Experimental results on the BraTS 2018, BraTS 2019, and BraTS 2020 datasets demonstrate that RLU-Net-GD outperforms state-of-the-art methods, significantly enhancing brain tumor segmentation accuracy. RLU-Net-GD utilizes residual connections in the encoder to enhance feature propagation and mitigate gradient vanishing, while deep supervision in the guided decoder, employing a weighted guided loss, ensures the extraction of salient features. These innovations enable the model to address challenges such as class imbalance and poor contrast, resulting in superior segmentation performance compared to state-of-the-art methods.
Diabetes is a widespread health issue that requires regular monitoring and professional attention for proper control. Combining IoT technology with traditional healthcare systems has enhanced the efficiency and quality of medical care. The sensor data is sensitive and may contain highly confidential information, such as medical diagnoses, clinical records, vital signs, and patient health data. Various privacy-preserving methods are currently used in disease prediction systems. However, patients may still be at risk of multiple health conditions. This study proposes a blockchain-integrated deep learning framework to enable secure data transmission within an IoT-based healthcare framework. Firstly, the medical data is gathered from IoT sensors, and then encrypted using the Sand Cat Optimized Elliptic Curve Cryptography (SOECC), where the optimal key is generated using Gaussian Perturbation and Brownian motion centered Sand Cat Optimization (GBSCO). Then, the encrypted data is stored in the blockchain, and the user decrypts it using the optimal private key. The disease prediction stage is implemented, where preprocessing is first performed to improve the dataset quality. After that, the dimensionality of features is reduced by using linear discriminant analysis (LDA). Lastly, classification is performed using the Optimized bidirectional gated recurrent unit (OBGRU). The outcomes demonstrate that the proposed framework achieves faster encryption and decryption processes, accurately determines a patient’s health condition with 99.15
Industrial and reliability optimization problems often involve complex constraints and discrete design choices, requiring the efficient and robust production of high-quality feasible solutions. This paper presents an advanced parameter reconfiguration-based Artificial Electric Field Algorithm (AI-AEFA) designed for constrained engineering design and reliability through a log-sigmoid-based parameter adaptation strategy and a chaotic-mapping behavior via a log-sigmoid-based parameter adaptation strategy and a chaotic mapping mechanism. We validate AI-AEFA on twenty-eight IEEE CEC 2017 constrained benchmark problems, fifteen large-scale industrial optimization problems, and seven reliability-redundancy allocation problems. Across these test suites, AI-AEFA demonstrates competitive or superior performance in terms of feasibility reliability and feasibility-aware solution quality (objective values reported under the adopted feasibility rule), while maintaining stable behavior across repeated runs. In addition, we integrate Shapley Additive Explanations (SHAP) to enhance interpretability by quantifying the influence of key internal factors (e.g., Coulomb’s constant, charge, acceleration, and electrostatic force) on the search dynamics. Overall, the results indicate that AI-AEFA is a robust, scalable, and interpretable optimization approach for practical constrained optimization applications.
Multivariate Time Series Anomaly Detection (MTSAD) is critical to industrial monitoring and system reliability. Despite numerous algorithmic advances, the field lacks dedicated sequence decomposition architectures for anomaly detection and effective mechanisms for fusing time-frequency features–limiting the performance of unsupervised MTSAD on spatiotemporally complex data, especially where labeled anomalies are scarce or unreliable. To address these limitations, we propose Time-Frequency Hybrid Decomposition (TFHD), a novel unsupervised framework for multivariate time series anomaly detection, built upon an encoder-decoder architecture. In the encoder, we introduce the Global Time-Frequency feature Fusion (GTFF) module, which integrates complementary information from both domains. GTFF comprises two components: a global time-domain feature extractor based on an external memory matrix, and a frequency-domain feature extractor leveraging the Fourier Transform. By jointly modeling decomposed subsequences from both domains, GTFF overcomes the representational limitations inherent in single-domain feature extraction. In the decoder, we present a Dual-Branch Feature Decoding (DBFD) mechanism, consisting of parallel decoding pathways for seasonal and trend components. This design enables the simultaneous reconstruction and refinement of distinct temporal patterns, effectively addressing the limitation of conventional single-branch architectures in capturing features from multiple decomposed subsequences concurrently. We evaluate TFHD on multiple benchmark datasets against state-of-the-art MTSAD methods. Experimental results demonstrate its superior performance in AUC-ROC, AUC-PR, and Affiliated F1-score (Genesis: 0.909; ASD: 0.832; CalIt2: 0.816), underscoring its efficacy in capturing complex anomalies.
Implementing agile software development(ASD) in Global Software Development (GSD) is challenging due to risk factors leading to loss of control, low product quality, and project failure. GSD firms continually enhance their process management to mitigate these risks and deliver quality products to clients. This study aims to develop a model to identify the key attributes for assessing agile project success. In addition, it proposes a cost-efficient effort-based prediction model to increase the likelihood of agile project success in a globally distributed environment. We evaluate project outcomes based on both cost and success probability. Harris Hawk Optimization (HHO) is employed to optimize the probability of agile project success with respect to cost, using efficiency as the fitness function. The Naïve Bayes Classifier (NBC) and Logistic Regression (LR) models are used to evaluate the proposed optimization framework. Data for training and testing the models were collected through a survey. The efficiency scores for the best solutions were found 0.5311 for NBC and 0.5850 for LR. The success probability improves for both client and vendor perspectives(approximately 60.4 p = 0.01 < 0.05 . The identified agile project attributes can assist Global GSD organizations in effectively implementing agile projects. The proposed prediction model focuses on ten agile attributes and their associated implementation costs. This model is critical for achieving successful and cost-effective agile project execution from both client and vendor perspectives.
The optimization of controllability in complex networks, particularly temporal ones, has been a significant focus in recent research, often leveraging metaheuristic approaches. This paper introduces a novel population-based metaheuristic approach to enhance the efficiency of controllability processes in temporal networks by leveraging a Temporal Max-Min Ant System (TMAS). The problem of identifying the Minimum set of Driver Nodes (MDS) for structural controllability in temporal networks is known to be NP-hard, rendering exact methods computationally infeasible for large-scale instances. Unlike prior methods that rely on greedy heuristics or sequential metaheuristics, our proposed method combines the explorative power of a parallelized TMAS (PTMAS) with an exact maximum flow algorithm to efficiently identify near-optimal driver node sets. The integration of strict pheromone bounding mechanisms prevents premature convergence and promotes robust exploration of the solution space. Experimental evaluations on multiple real-world temporal networks demonstrate that our method significantly outperforms existing conventional methods, reducing the number of required driver nodes by up to 20
Facial expression recognition (FER) remains a challenging problem in computer vision due to variations in illumination, pose, and occlusion under complex real-world conditions. To address these issues, this paper presents a Multiscale and Permute Attention Network (MPA-Net) that jointly models global, local, and salient features for robust emotion recognition. Specifically, a pre-extraction block captures low-level visual patterns, while a multiscale block enhances feature representation by integrating information from multiple receptive fields, improving robustness to occlusion and pose changes. Furthermore, a space-wise split block enables the network to emphasize discriminative local regions, and a permute attention mechanism establishes interactions among channel, spatial, and height dimensions to achieve comprehensive feature fusion. Extensive experiments on three benchmark datasets-FER2013, RAF-DB, and AffectNet-demonstrate that MPA-Net achieves competitive performance (72.28
High-voltage insulators must be highly reliable because failures can cause serious operational problems and safety risks in power transmission networks. This study presents a classification framework for 15 kV insulator degradation utilizing a metaheuristic optimization algorithm and a BiGRU-CNN deep learning model. The suggested hybrid methodology, which combines the bidirectional gated recurrent unit (BiGRU) and the convolutional neural network (CNN) to efficiently capture the temporal dependencies from environmental data, performs better and is more accurate when the hyperparameters are optimized using the Improved Particle Swarm Optimization (IPSO) methodology. Experimental analysis was conducted on a comprehensive dataset of environmental and operational parameters collected from 15 kV polymeric insulators under real field conditions. Comprehensive analysis is conducted to compare the performance of the proposed IPSO-BiGRU-CNNs with other state-of-the-art deep learning techniques, including RNN, CNN, GRU, LSTM, BiRNN, BiGRU, and CNN-BiGRU models by various evaluation metrics. The model’s performance is further improved by integrating IPSO algorithms for hyperparameter tuning. The most notable improvements are 23.28
Federated Learning (FL) has emerged as a decentralized alternative for Human Activity Recognition (HAR), allowing models to be trained across devices without compromising user privacy. However, existing FL frameworks still require a large amount of labeled data to achieve effective performance, particularly when facing significant data heterogeneity challenges in natural environments. To address this, we propose Federated Active User Participation (FAUP), an adaptive FL framework that reduces data requirements while maintaining a low communication overhead. FAUP employs adaptive clustering to prioritize “Active Users” with higher signal variance, enabling model aggregation based on similar activity patterns. By leveraging these informative data samples, FAUP accelerates convergence while maintaining high accuracy. We evaluated FAUP on four publicly available HAR datasets involving 148 users. The results demonstrate that FAUP achieves comparable accuracy to existing frameworks while requiring only 50
Reliable fall detection is essential for ensuring the safety of older adults in smart healthcare environments. Existing vision-based deep-learning fall detection systems often suffer from two major issues: high computational cost and susceptibility to false detections across varied environments. To address these problems, we propose 3D-LiteNet, a novel lightweight 3D convolutional neural network (CNN) model derived from 3D MobilenetV2, specifically optimized for fall detection in video sequences. The innovation of the proposed approach lies in its ability to reduce model parameters by over 91
ColoRectal Cancer (CRC) is an increasing worldwide health concern, marked by persistently high mortality rates. Early detection is crucial for increasing life expectancy and reducing the risk of illness. Conventional visual inspection of biopsy slides stained with haematoxylin and eosin is time-consuming and subjective. Microscopic investigation or naked-eye inspection by clinical experts is always suboptimal. Moreover, this approach requires extensive and high-cost genetic testing. Offering a new paradigm, with a recent advancement of image processing and machine learning techniques, we propose a robust CRC subtype classification framework called “DC3NET” (abbreviated as Depthwise separable Convolutional Colorectal Cancer Neural Network). DiSC (Depthwise Separable Convolutional Neural Network) is used for extracting hidden morphological features from histopathological images. Then, the multiclass classification is performed to distinguish the CRC subtypes using an Optimized Support Vector Machine. Meta-visualisation techniques such as GradCAM, occlusion sensitivity, gradient attribution are also integrated to improve clinical interpretability. The evaluation metrics such as sensitivity, accuracy, precision, F1 score are evaluated. Our proposed DC3NET model resulted in an impressive validation accuracy of 94.92
This study develops an efficient prediction framework for urban fire evolution by integrating GIS-based modeling with a hybrid physical–empirical fire model. The framework significantly reduces computational workload and simulation complexity while maintaining physical interpretability. Validation against real fire incidents confirms its ability to accurately reproduce key fire propagation dynamics, both qualitatively and quantitatively. Simulation results further highlight the critical influence of ignition location and wind conditions on urban fire spread, supporting the formulation of targeted prevention strategies and emergency plans. Notably, the framework boasts high computational efficiency, as it can simulate 48-hour fire scenarios involving thousands of buildings in less than one minute on a standard personal computer. This study provides a practical and scalable tool for urban fire risk assessment, real-time prediction, and resilient urban design.
