Journal of Intelligent Systems and Internet of Things

Automated vehicle monitoring in intelligent transportation systems must operate reliably around the clock, including under conditions that routinely cripple conventional visible-light cameras: night, glare, shadows, and adverse weather. This paper proposes a modular Internet of Things (IoT) architecture for thermal-based vehicle detection, classification, and trajectory analysis, together with a four-phase deployment roadmap that connects public-dataset evaluation to live-traffic field validation. The system integrates longwave infrared (LWIR) imaging (8–14 𝜇m) with YOLO-family deep learning detectors (YOLOv8/v11/v12) and multi-object tracking algorithms (ByteTrack, BoTSORT, StrongSORT), deployed across NVIDIA Jetson edge devices and cloud infrastructure through JSON/MQTT formalized data contracts. The primary novel contribution is a system-level integration framework that bridges the gap between component-level algorithmic research and operational deployment. Concretely, this work: (i) defines five functionally independent modules with explicit interface specifications and latency budgets not previously formalized in the thermal-ITS literature; (ii) introduces quantified decision gates linking progression criteria directly to published benchmark values; (iii) provides region-specific operational availability estimates derived from empirical weather-degradation data; and (iv) integrates domain adaptation, GDPR compliance, edge hardware budgets, and regulatory WIM frameworks within a single coherent system blueprint. Domain adaptation strategies reported in peer-reviewed literature recover 20–50% of cross-dataset mAP degradation (typically 10–30%) caused by sensor and scene variability; these figures are literature benchmarks, not results obtained in this work. An optional weight-estimation module (Module 4) based on recent vision-based and bridge WIM validation studies is treated as an exploratory extension requiring site-specific validation.

In the intelligent teaching environment, it is indirect and difficult for teachers to capture learners’ learning attitudes and behaviors through digital learning behavior data provided by intelligent platforms. The purpose of this paper is to improve the precision of student behavior detection in teaching, and to provide teachers with a more reliable basis for making teaching plans. The Yolov8 algorithm is applied to student behavior recognition, and a bounding box loss function based on dynamic focusing mechanism is introduced to make a balance between samples with good regression quality and poor regression quality. Through experimental analysis, we can see that the real-time behavior tracking algorithm combined with Yolov8 proposed in this paper has a good application effect in student behavior detection. Moreover, it not only improves the precision of student behavior recognition, but also improves the stability of the algorithm, which is conducive to the effective development of subsequent smart teaching models.

One of the most dangerous and deadly illnesses that people can face in their lives is cancer. Among all cancers, skin cancer is one of the most damaging, hazardous, and potentially fatal to a person's life. If not detected it and treated initially, it will extend to other body parts soon and lead to the deadliest situation.  It will spread quickly when the skin tissue areas are exposed to sunlight, mostly because skin cells in the designated location develop quickly. An automated skin tumor recognition system is the main requirement in order to detect skin cancer early, minimize time and effort, and save human lives. The most popular and successful methods for classifying skin cancer are the techniques of image processing and deep learning models. So, there is a need for an automated healthcare system to detect and classify skin lesions. We proposed a CNN model for classifying skin tumor images in our work. We have trained CNN models like AlexNet, VGG16, ResNet50, and Inceptionv3 using transfer learning techniques and observed the performance accuracies of all the models. The dataset used in our work contains two types of benign and melanoma skin tumor images, which are classified into two kinds through the Convolution Neural Network models. We used preprocessing techniques to clean our data, and data augmentation was also used to generate more data. As we know, deep learning models need more data to train and test the models. In all our model implementations, we have used all the features from the image while training the models for classification. Finally, we used the transfer learning techniques in our implementation models to improve the accuracy of each Image classification model. We trained the three models with different optimizers: Adam, Adadelta, and SGD. The proposed model (Modified AlexNet) provides better results, with approximately 96.75% for Training accuracy, 94.43% for Validation accuracy, and 94.11% for Testing Accuracy. The proposed model's performance results are compared with the state-of-the-art models like AlexNet, InceptionV3, VGG16, and ResNet50.

Ecommerce Platforms specifically in Retail domain be it a brick and morter store or an online shopping application has enormous user data from the behavioral, click stream, page visits, abandoned carts, user think time or dwell time. And from the retail stores where the data captured from Internet of Things (IoT) with respect to the shelve movements, visitor counts, IoT signals arising from RFID tags, beacons, smart sensors, proximity to specific products, kiosk interactions, self-checkout kiosk provide enormous data for hyper personalization. Traditional Singular Value Decomposition (SVD) algorithms suffer with the data sparsity and computational complexity when fed with such large data. Also the SVD relies on the historical patterns to find latent features which may not be very much helpful for the cold start personalization. Consumer behaviors and patterns are non-linear, for ex- ample time spent near a shelf in a Retail Store or the time spent on a categories page in online application and with the filters of the categories. SVD might capture these main trends but will miss subtle high frequency signals that drive the hyper personalization. To overcome this problem, the proposed research employs a significant latent core factor SVD. The proposed technique includes decomposing a large and sparse matrix that captures real-time interactions between users and products into matrices that permit the proposed model to forecast personalized product recommendations based on existing data. Large Language Models (LLM) were used to improve the process of feature extraction post the data imputation after the initial data preprocessing. The proposed research employs the Amazon product review dataset to evaluate the proposed significant latent core SVD. When compared to traditional SVD and state-of-the-art methods such as LightGCN and BERT4Rec, the proposed significant latent core factor SVD achieves lower error rates.

Earthquake prediction is one of the most challenging problems in geophysical science, and conventional approaches have proven arduous in capturing the complexity and non-linearity of seismic measurements. The multidimensional nature of earthquake variability, along with class imbalance and the strong dependence of prediction results on hyperparameters, necessitates the development of more robust and flexible predictive models. In this paper, we introduce a bio-inspired ensemble learning method based on the Somersaulting Spider Optimizer (SSO) for dynamically adjusting classifier weights in earthquake classification. The proposed method addresses limitations of existing weighting strategies, which primarily focus on maximizing classifier contribution based on performance characteristics. Experiments were conducted on an earthquake dataset augmented with features modeled and mapped by time, space, and magnitude to capture patterns of seismic events. We compared the SSO-optimized ensemble with BaggingClassifier, CatBoost, HistGradientBoosting, LightGBM, and DecisionTree, as well as traditional ensemble approaches. Results show that the SSO-boosted ensemble achieved superior performance, with an accuracy of 97.01%, sensitivity of 97.04%, specificity of 99.36%, precision of 97.64%, and an F1-score of 97.33%, outperforming other models and traditional ensembles. These improvements were confirmed statistically using Wilcoxon signed-rank tests, while visual analyses demonstrated enhanced stability and generalization. Overall, the integration of bio-inspired optimization and ensemble learning shows strong potential to overcome challenges in earthquake forecasting and to support reliable early warning and disaster preparedness systems.

Human-computer interaction (HCI), artificial intelligence (AI), and HI are in high demand these days. In fields like marketing, client feedback analysis, security, and healthcare, facial expression- grounded emotion recognition becomes a pivotal tool for comprehending mortal feelings. Facial expressions like fear, disgust, surprise, anger, sadness, and happiness are pivotal pointers of emotional countries. Businesses can ameliorate client gests by relating these pointers and measuring client satisfaction with goods or services. The discovery of mortal feelings has been achieved with machine literacy algorithms like support vector machines and arbitrary timbers. The effectiveness of deep literacy models for emotion discovery has been validated by earlier studies that employed Convolutional Neural Networks (CNNs) to reliably classify feelings grounded on facial expressions. Likewise, recent developments in deep literacy, particularly the operation of Convolutional Neural Networks (CNNs), have significantly increased the delicacy of facial emotion recognition and interpretation from images and live camera aqueducts. In order to reuse face images with CNN models for real- time emotion recognition, our exploration attempts to produce an emotion recognition system using Python and OpenCV. The current study describes how to watch live videotape aqueducts for facial expressions to identify which of the seven linked feelings is most likely to do. This system provides emotional behavior in real time when needed.

Recently, Deep learning (DL) models are increasingly used in Test Case Prioritization (TCP) tasks combining partial and imperfect test case (TC) information into accurate prediction models. Various DL algorithms have been created to improve TC failure prediction and prioritization in CI settings. Among them, Deep Reinforcement Prioritizer (DeepRP) model is developed using Deep Reinforcement Learning (DRL) and Deep Neural Network (DNN) for efficient TCP on huge test suites. But, the model's labelling task is interrupted early, creating difficulty in learning TC features for unlabeled training TCs due to limited resources. To solve this, Deep Graph Reinforcement Prioritizer (DeepGRP) is proposed in this paper to learn the TC features from unlabeled training data for efficient TCP in Regression Testing (RT). In this method, graph neuron stimulation attributes for TCs are created to retrieve the activation graph across DNN layers of DeepRP. The connectivity neuron link defines the activation graph. The proposed deep graph (DG) recognizes the DNN neurons as nodes and the adjacency matrix as the connectivity link among the nodes. Also, the message passing mechanism is applied to aggregate the structural information from the adjacency matrix with neighbouring node features to enhance TCP. By applying this mechanism, DeepGRP captures the high-order dependencies among neurons for efficient activation features which overcomes the traditional activation models and improves the TCP at large scale RT.  The DG model prioritizes TCs using Learning-to-Rank (L2R) which learns node attributes from TCs. This enables for better DNN testing efficiency by detecting vulnerabilities early and lower development time for efficient TCP and tackling the difficulty of learning TC characteristics for efficient TCP. Finally, the testing findings suggest that the DeepRP can improve the TCP for large TSs when compared to other common algorithms.

Microorganisms are commonly found in our daily living environments and play a crucial role in environmental pollution control, disease prevention, and treatment, as well as food and drug production. To fully utilize the diverse functions of microorganisms, their analysis is essential using Intelligent Systems. Traditional analysis methods can be labor- intensive and time-consuming. As a result, image analysis using Intelligent Systems i.e. machine learning or deep learning have been introduced to improve efficiency. Deep learning networks algorithms such as CNN contain a stack of multi-layer, the first layer and the last are the input and output layers, between them are the hidden layers to extract and learn many features in images, recurrent network algorithms (RNN) combined with convolution neural network (CNN), these networks allow to process a series of images to extract the crucial information from images and also these algorithms help to minimize the size of images and reduce the redundancy in microrganisms images According to previous studies, these algorithms are the most used to classify the images of microorganisms. However, the classification of microorganism images presents several challenges these include the need for robust algorithms due to varying application contexts, the presence of insignificant features, along various analysis tasks that need to be addressed. The research summarizes significant advancements that tackle these challenges through deep learning and machine learning methods. Current obstacles, gaps in knowledge, unresolved issues, limitations, and difficulties in classification techniques are also discussed.

The proposed study analyses a hybrid deep learning method to monitor a vehicle with drones with augmented simulated GPS data to increase awareness and localization accuracy. The system combines both the high detection speed of a real-time YOLOv5 with the high recognition accuracy of task-driven Faster R-CNN, which makes the performance of the system quite balanced, fully applicable to the application of aerial surveillance enforcement. The results will mimic realistic monitoring conditions since synthetic aerial scenes were produced in which vehicle density is randomly distributed and simulated geolocation data. Both models were applied in the processing of each scene and the resultant images were combined by a voting scheme. The hybrid system had an accuracy of 1.00, recalls 0.90, and F1 score of 0.95- it performed higher than the Faster R-CNN alone (F1 score:0.89) and higher in different conditions. The novelty of the proposed research is based on the fact that the invention combines the methods of dual-modality object detection (visual + spatial) and the use of a GPS base, which allows not only visual object detection but also object positioning. As opposed to the approaches previously used, based on single-modality models and without consideration of the data on geolocation, the framework achieves the integration of object recognition and useful mapping. The suggested system is lighttrack, economically feasible, and it is conveniently deployable to present scalable real-time traffic tracking, smart city planning, and aerial autonomy surveillance.

The promoter is the part of DNA, which is responsible of initiating RNA polymerase transcription of a gene. The location of this part of DNA is upstream the transcription start site. According to researches, the genetic promotors contribute majorly in many human diseases such as cancer, diabetes and Huntington’s disease. Therefore, promotor detection corresponds as a very crucial task. In this study, a hypered detection system, which integrates biologically developed feature extraction with traditional machine learning (ML) algorithms in addition to use Long Short-Term Memory (LSTM) network as a deep learning approach, has been proposed. The dataset used includes 106 nucleotide sequences. Results obtained from the study show that the perfect performance across all metrics (accuracy, sensitivity, specificity, precision, and F1-score) has been achieved when Naive Bayes used as a classifier, which reach 100% and AUC=1.The confusion matrix analyses and ROC curve confirm that LSTM model achieved 100% training accuracy and 84.38% test accuracy. The architecture and performance of the proposed model make it applicable in IoT-based intelligent genomic and healthcare systems, which enabling real-time and remote promoter detection.

Sign language (SL) detection and classification for deaf persons is an essential application of machine learning (ML) and computer vision (CV) techniques. It covers emerging forms, which acquire SL implemented by entities and convert them into auditory or textual output. It is highly significant to understand that determining a correct and robust SL detection approach is a very challenging due to many tasks such as alterations in occlusions, and lighting states in hand actions and forms. Consequently, the CV and ML models is must for testing and training. A Hand gesture detection method discovers beneficial for hearing and speaking-impaired individuals by creating usage of convolutional neural network (CNN) and human-computer interface (HCI) for classifying the constant signals of SL. In this article, an Improved Fennec Fox Algorithm for Deep Learning-Based Sign Language Recognition in Hearing and Speaking Impaired People (IFFADL-SLRHSIP) technique is proposed. The presented IFFADL-SLRHSIP technique main intention is to provide effectual communication between deaf and dumb persons and normal persons utilizing CV and artificial intelligence techniques. In the IFFADL-SLRHSIP model, the enhanced SqueezeNet model is used to capture the intricate patterns and nuances of SL gestures. For detection of the SL classification process, the recurrent neural network (RNN) method is used. To optimize model performance, the improved fennec fox algorithm (IFFA) is applied for parameter tuning, enhancing the model's precision and efficiency. The experimental outputs of the IFFADL-SLRHSIP algorithm are legalized on the SL dataset. The simulation outcomes demonstrate the greater outcomes of the IFFADL-SLRHSIP approach in terms of diverse measures.

Blind and visually challenged people face the range of practical issues by undertaking outside travels as pedestrians. In the last decade, various beneficial devices is investigated and established to assist people with disabilities move independently and safely. Anomaly detection in pedestrian paths for visually impaired individuals, using remote sensing (RS), is crucial for improving pedestrian traffic flow and safety. Engineers and investigators can create efficient methods and tools with the effect of computer vision (CV) and machine learning (ML) to recognize anomalies and alleviate possible security hazards in pedestrian walkways. With recent progress in deep learning (DL) and ML fields, researchers have realised that the image recognition problem is supposed to be developed as classification problems. This paper proposes a Coati Optimization Algorithm-Based Parameter Tuning for Pedestrian Walkways with Transfer Learning Model (COAPT-PWTLM) technique. The main goal of COAPT-PWTLM technique is to provide automatic detection of pedestrian walkways for disability using advanced models. Initially, the median filtering (MF) is employed in the image pre-processing stage to eliminate the noise from an input image data. Furthermore, the SquezeNet1.1 model is utilized for feature extraction. For the classification process, the multi-layer autoencoder (MLAE) model is implemented. Finally, the modified update coati optimization algorithm (MUCOA) model adjusts the hyperparameter range of MLAE method optimally and results in improved classification performance. The experimental validation of the COAPT-PWTLM is verified on a benchmark image dataset and the outcomes are evaluated under dissimilar measures. The experimental outcome underlined the progress of the COAPT-PWTLM model over the existing models.

In the image-captioning field, the excellence of produced captions is vital for the effectual interaction of visual content. Image Captioning is the main task, which unites computer vision (CV) and natural language processing (NLP), where it goals to produce graphic legends for images. A dual-fold procedure depends on precise image perception and alters language understanding both semantically and syntactically. It is gradually challenging to stay up with the modern study and consequences in image captioning owing to the developing amount of knowledge accessible on the topic.  This analysis examines into deep learning (DL) to tackle the tasks challenged by individuals with graphic impairments, targeting to improve their visual insight via advanced technologies. By tradition, the visually impaired have trusted physical support and adaptive helps for understanding and navigating visual content. With the beginning of DL, there is a unique chance to develop this scenery. In this paper, we offer an Advanced Deep Learning Method for Image Captioning Based Using Customized Transformer with a Global Optimization Algorithm (ADLIC-CTGOA). The foremost aim of ADLIC-CTGOA model is to focus on the initiation of the effectual textual image captioning of an input image. Initially, the ADLIC-CTGOA method employs preprocessing phase to enhances both image and text data: images undergo noise removal and contrast enhancement to improve quality, while text is processed by removing numbers, converting to lowercase, and text vectorization. Next, the customized swin transformer is employed for feature extraction to capture fine-grained visual features from images. In addition, the BERT Transformer model is deployed for image captioning process. To enhance the performance of proposed technique, the chaotic Aquila optimization (CAO) technique was applied for parameter tuning for enhancing the performance. A wide sort of simulation studies are executed to ensure the improved performance of ADLIC-CTGOA system. The comparative result exploration reported the betterment of the ADLIC-CTGOA model on recent approaches in terms of different evaluation measures.

Driver drowsiness detection is an important aspect of intelligent transportation systems that aim to reduce fatigue-related accidents. The existing schemes based on threshold-based method, or deep-learning based models often found to be associated with issues in terms of flexibility, computational efficiency, or capacity for real-time performance. This paper presents a development of two-stage hybrid framework for driver drowsy detection, where the first stage utilizes a fuzzy-logic based approach applied to physiological measures, facial feature, head position, blink duration, and eye movements to produce lightweight and adaptive analyses of sleepiness in drivers. The second stage consists of a hybrid quantum-classical neural network (HQCNN), in which convolutional neural networks (CNN) extract spatial features whereas quantum fully connected (QFC) components apply entanglement-based transformations to improve both feature characterization and classification accuracy. The experimental result validates effectiveness of the proposed hybrid method with 94% accuracy, and better than traditional CNNs with real-time capability. The proposed framework is developed to achieve a balance between computational efficiency and classification/decision quality thereby making it suitable for driver monitoring in real-time application.

Natural Language Processing (NLP) and Network Science were combined to study emotional contagion dynamics in social media networks. We simulated the diffusion of emotions through users on a synthetic interaction network using sentiment-labeled Twitter data and a graph-based model. We explored the relationship between graph metrics, including centrality and clustering coefficient, on emotion propagation and stability. The findings show that emotion intensity converges through the network and that both weak coupling of central nodes and moderate cluster structures dampen the spread of emotion. A community-level analysis reveals more alternative results, such as the fact that emotions differ in polarity between communities. Our work demonstrates a better understanding of how emotional behavior in online environments can be adjusted using semantic measures, which offer a means to characterize the relevance of information online and the interconnected relationships among emotionality.