In this study, the hyperparameters of Stochastic Gradient Boosting, LightGBM, and Histogram-based Gradient Boosting models were optimized using evolutionary algorithms. The main goal of this research is to find the most effective combination of hyperparameters for each model. Their goal is to increase accuracy and improve computational efficiency. The study was conducted on the basis of real clinical data. The data were obtained from the Samarkand City Endocrinology Center, using a dataset of diabetes-related and clinical indicators. The data were initially cleaned and processed using normalization, imputation, and 5-point cross-validation. We used five evolutionary strategies to fine-tune the main hyperparameters: Differential Evolution (DE), Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Covariance Matrix Adaptation Evolution Strategy (CMA-ES), and Dynamic Ensemble with Hyperband (DEHB).In the optimization, the F1-macro was chosen as the main fitness function. This allows for a balanced accuracy assessment for multi-class classification. Across all three prototypes, evolutionary tuning consistently led to better results than Grid Search and Random Search. SGB+DE accuracy 0.9098, F1-macro 0.9096 gave the best results for the single model, whereas DEHB and CMA-ES worked better for LightGBM and HGB, respectively. The absolute increases above the unadjusted baselines ranged from 0.3 to 0.6 percentage points. They were small but could be repeated reliably across all evaluation criteria. During this study, the models were evaluated on the metrics of accuracy, precision, recall, and F1-macro.

Bangla Sign Language is a unique sign language. Due to a lack of interpreters, the hearing- and speech- impaired community face challenges while communicating with the broader community. Recent studies have been con- ducted to reduce the gap between these two communities. But most of the researchers used a dataset with a controlled environment. We know the performance of a system highly depends on dataset quality. In this paper, we have created a new dataset, “BanglaSignSet” including 46 unique signs with over 10k images. We have carefully annotated and labeled the images using Roboflow. Our proposed dataset, “BanglaSignSet” consists of images with high resolution, good qual- ity, and adequate variation in environment and person. The constructed dataset has been trained using the most recent deep learning model, such as YOLOv8. We have also implemented different versions of the YOLOv8 model, such as YOLOv8n, YOLOv8s, and YOLOv8m. Additionally, we evaluated EfficientNet-B0 as a classification-based baseline to broaden the experimental comparison. The performance of models has been measured using different evaluation metrics such as mAP, precision, recall, and f1 score. A comparative analysis has been conducted based on the performance of the model. By comparative analysis we found a well-suited model, YOLOv8n, to deploy into a web-based application. To find the suitable model to deploy, we have considered factors such as memory requirement and inference speed. We have integrated the YOLOv8n model into a web application using the Python language. We have also tested the web application on Android devices and laptops. The web application detects signs from image input successfully.

Image generation methods, including inpainting, are evolving rapidly; however, high memory requirements continue to limit their practical deployment. As a result, the efficient utilization of Latent Diffusion Models on edge devices has become increasingly important. This work explores techniques for reducing memory usage in Latent Diffusion Models while preserving their generative capabilities.

We propose a resource-efficient inpainting method optimized for CPU-based inference, based on a combination of VAE tiling, attention slicing, and dynamic region-of-interest slicing. Experimental results demonstrate that the model's memory footprint can be significantly reduced while maintaining output quality, without substantial increases in computation time, enabling execution on systems with as little as 4 GB of memory and only two processing cores. While the introduced optimizations, particularly those based on localized image processing, introduce an inherent trade-off between memory usage and computational cost, resulting in longer inference times compared to GPU-accelerated solutions, they demonstrate strong potential for deployment in memory-limited environments.

Additionally, we provide analysis of key deployment bottlenecks, including model compilation for cold-start overhead mitigation, proper runtime configuration and scheduler selection. These findings confirm the feasibility of effectively deploying Latent Diffusion Models for inpainting tasks on CPU-only, resource-constrained platforms, thereby broadening their applicability to edge computing scenarios.

Accurately predicting corrosion inhibition efficiency (IE) remains a significant challenge in electrochemical systems because inhibitor performance depends on a complex interaction between molecular electronic structure, solution chemistry, and adsorption thermodynamics at the metal–electrolyte interface. Although machine learning (ML) models have shown strong predictive potential in corrosion studies, most existing approaches function as purely data-driven regressors and do not explicitly incorporate adsorption-based electrochemical principles. As a result, they may generate thermodynamically inconsistent predictions, particularly when applied beyond the range of the training data. In this work, a physics-informed hybrid machine learning framework is developed for adsorption-controlled corrosion systems. A Physics-Informed Neural Network (PINN) is used to incorporate the Langmuir adsorption relationship directly into the training process, ensuring monotonic consistency between surface coverage (θ), adsorption equilibrium constant (K_ads), and inhibition efficiency. To capture nonlinear interactions and multivariate effects beyond the analytical adsorption model, a Gradient Boosting Regressor (GBR) is introduced as a complementary data-driven component. The predictions from both models are then integrated using Ridge-based stacking, allowing a balanced combination of physical interpretability and statistical flexibility. The framework is trained using a composite dataset consisting of 100 experimentally reported corrosion inhibitors and 1000 physics-consistent synthetic samples generated within experimentally observed parameter ranges. Across five-fold cross-validation, the hybrid model achieves stable predictive performance with a mean test-set coefficient of determination of R² ≈ 0.90 while preserving adsorption-consistent behavior. The standalone GBR provides the highest numerical accuracy, while the PINN improves physical consistency and interpretability; the stacked model combines these advantages effectively. Overall, the results show that embedding adsorption thermodynamics within machine learning models improves predictive reliability, physical consistency, and model transparency in corrosion inhibition studies. Rather than introducing a fundamentally new algorithm, this work demonstrates a corrosion-specific integration of physics-informed learning and hybrid regression for more reliable inhibitor screening. The proposed framework provides a structured and reproducible approach for early-stage corrosion inhibitor design and can be extended to other adsorption-governed systems, including catalysis, surface protection, and electrochemical materials engineering.

The results of studies on an innovative technology for complex presowing seed treatment, which includes the application of protective and nutrient substances with subsequent drying, are presented. Traditional apparatuses with mechanical agitators or rotating bottoms have substantial limitations, such as a high risk of mechanical damage to seeds, the formation of agglomerates, and a low intensity of heat and mass transfer. To solve these problems, the application of an inhomogeneous jet-pulsating fluidization in a self-oscillating mode is put forward as a viable solution. A physical model of the interaction of a gas coolant with granular material of non-spherical (ellipsoidal) shape is theoretically and experimentally validated using wheat grains as an example. It has been experimentally proven that in a gas flow, seeds orient themselves with minimal projection along the flow direction. This reduces hydrodynamic resistance and leads to a local increase in bed porosity to 0.5-0.9. It was established that the introduction of a coolant forms an intense asymmetric spouting motion. At the same time, cyclic entrainment of about 40 % of the bed mass into the freeboard space occurs with a frequency of more than 1.5 hertz. The ratio of the gas bubble volume to the initial bed volume increases to 37 %. This specific hydrodynamic condition provides active volumetric mixing with a significant increase in the interfacial contact area and intensification of heat and mass transfer processes at low temperatures (not exceeding 40 °C), which significantly mitigates the risk of thermal degradation and mechanical impact to the seeds compared to traditional methods.

Vulnerability detection is a preventive approach for performing rigorous maintenance on software projects. In cross-domain settings, in-domain methods cannot achieve high VD performance due to differences in data distribution and labeling. Existing cross-domain VD methods suffer from the following limitations: 1) They require matrix trans-formation to meet sequence embedding criteria, 2) Feature matching relies on effort-intensive graph-based analysis that results in high computational cost, 3) Each cross-domain solver is generally designed for a specific programming lan-guage, preventing a global domain adapter. To address these problems, we present VulSD (Vulnerability detector using Static and Dynamic analysis), a cross-domain approach based on static code metrics and dependency analysis. Unlike existing methods, VulSD combines an embedding matrix produced by Word2Vec with static and dynamic code features. Additionally, VulSD employs Spearman analysis to convert constant features for compatibility with the training process. Finally, a deep learning model is established using the R deepnet library. VulSD achieves an average accuracy of 84.2% on large benchmark datasets (DiverseVul, Devign) and 70-77% on real-world project datasets. Performance varies across targets, with best results on C/C++ benchmarks and more modest gains on mixed-language and smaller project datasets.

This paper focuses on real-time anomaly detection in surveillance video using YOLOv8, the latest in the YOLO object detection series, integrated with spatio-temporal analysis. The system aims to detect abnormal behavior in crowded environments by combining spatial object detection with temporal activity analysis. YOLOv8 is used to detect and track individuals in video frames, while a 3D Convolutional Neural Network (3D CNN) processes sequences of frames to identify behavioral anomalies based on movement patterns. Three variants of YOLOv8—Nano (n), Small (s), and Medium (m)—are evaluated for performance trade-offs in accuracy, processing speed (FPS), and latency. Results show YOLOv8n offers the best real-time performance, while YOLOv8m provides higher accuracy at the cost of increased latency. The system uses the UCF-Crime dataset for training and testing, and metrics such as accuracy, FPS, and latency are used for evaluation. The modular pipeline supports scalability and real-time deployment, with visual outputs aiding interpretation. By integrating object detection with spatio-temporal modelling, the system effectively identifies anomalies such as loitering or sudden movements. Future work includes refining detection accuracy using labelled anomalies and exploring advanced models like Transformers for improved temporal understanding. The significance of this research lies in its ability to combine lightweight real-time object detection with effective temporal behavior modeling within a scalable and modular architecture. The proposed framework contributes to the advancement of intelligent surveillance systems by improving anomaly detection reliability while maintaining computational efficiency suitable for deployment in smart cities, public safety monitoring, and edge-based surveillance applications.

A Modified Group Method of Data Handling (MGMDH) is a component of a synthetic method of constructing multivariate polynomial regression given by a redundant representation. The MGMDH is used to construct multivariate linear regression given by a redundant representation in the case when a decomposition method, which is also a component of the synthetic method, allowed to estimate with a given accuracy the values of unknown coefficients for nonlinear terms of a multivariate polynomial regression. As statistical studies have shown, the MGMDH efficiently finds the correct structure of a multivariate linear regression when the realizations of a random factor in the tests are an order of magnitude smaller than the modules of the corresponding values of the regression to be determined. Only in this case, the use of the regularity criterion in the MGMDH almost always allows finding the correct structure of a multivariate linear regression given by a redundant representation. In this paper, the MGMDH is adapted for the case when during the tests the modules of the random factor implementations and the values of the regression to be determined take values of the same order, which significantly increases the efficiency of using the MGMDH for constructing multivariate linear regressions given by redundant representations. It is obvious that the adapted MGMDH for multivariate linear regressions given by redundant representations presented in this work is easily transformed using the standardization operation for the general problem of constructing multivariate regression given by a redundant representation in the case when the unknown coefficients are linear in the regression structure.

The article presents a model of dynamic processes occurring in non-isolated populations that differ in their habitat and mode of nutrition. The results of theoretical studies carried out on the basis of this model show the decisive influence of the ratio of the coefficients of inter-subpopulation competition on qualitative changes in the behavior of the system and individual subpopulations. This ratio is also the main factor influencing the formation of the dominant subpopulation in the system. It has been shown that the system-wide dynamics of subpopulation processes significantly depends on the reproductive potential of all subpopulations and on the mass fraction of individuals that, according to their phenotypic properties, are related to the parents. In this case, the mass fractions of individuals (transition coefficients) must correspond to the condition of closed system and be in specified intervals. It has been established that subpopulations in real life can exchange descendants, which, in turn, can significantly affect the numerical and qualitative aspects of the dynamics. Using the example of a two-dimensional system, the relationship between the sum of the main elements of the transition coefficient matrix and the mutual dependence of subpopulations, as well as their transition to qualitatively different levels, is shown. The bifurcation properties of the model of subpopulation dynamics with a Lotka–Voltaire type function in basic quality have been studied. An approximate justification of possible bifurcations of the system allows us to evaluate the factors that qualitatively influence the dynamics of the system and develop a number of recommendations to prevent the occurrence of catastrophes and collapses in the system.

Generating realistic images from textual descriptions remains a core challenge in artificial intelligence, with broad applications in assistive technology, virtual environments, and creative media. Existing text-to-image synthesis models often struggle with fine-grained semantic alignment and motion-aware scene generation, particularly in dynamic or complex prompts. This paper presents MoCoGAN+ATT, an enhanced framework that extends the MoCoGAN architecture by integrating attention mechanisms and Bidirectional Encoder Representations from Transformers (BERT) to extract and align rich semantic features from text. The attention module enables precise correspondence between textual concepts and visual components, leading to semantically faithful and visually coherent image generation. We evaluate MoCoGAN+ATT on five benchmark datasets—COCO, CUB-200-2011, Oxford-102 Flowers, MSR-VTT, and Visual Genome—demonstrating notable improvements over existing baselines. Specifically, on the COCO dataset, the proposed model achieved an Inception Score of 28.71, FID of 11.91, and R-Precision of 94.92; on CUB-200-2011, it obtained 27.36, 12.72, and 93.53 respectively; on Oxford-102 Flowers, the model achieved 28.63 (IS), 14.53 (FID), and 73.78 (R-Precision); on MSR-VTT, results were 28.01, 12.62, and 96.43; and on Visual Genome, we recorded 28.15, 17.93, and 94.52. The key novelty of this work lies in fusing motion-aware generative modeling with fine-grained attention-guided textual conditioning for dynamic image synthesis. These results highlight the effectiveness of combining attention-based textual conditioning with motion-aware generative modeling and point toward promising future directions for advancing multimodal image generation.

This article explores the use of modern technologies to train pre-service specialists to work productively in smart cities in Ukraine. Firstly, the concept and main components of digital infrastructure as a training foundation for future professionals were identified. Next, training technologies were determined using the expert evaluation method. Subsequently, a survey was conducted among students and lecturers at higher education institutions. The results showed the most important factors of professional training. After applying these, a model for integrating training into a smart environment was developed, along with recommendations for its use. The research found that effective professional training programs for future smart city development specialists require cooperation among municipalities, universities, businesses, civil society organizations, and residents. Following implementation of the research, we concluded that the most useful training technologies were Smart Ukraine, the Lviv IT Cluster, the IT Ukraine Association, and the Association for Innovative and Digital Education. A variety of teaching methods were employed in pilot programs to demonstrate improved student skills, including problem-based learning, interdisciplinary projects, and VR/AR simulations. The research also highlights the importance of ongoing professional development in fostering innovation readiness.

Plant disease detection is vital for agricultural sustainability and food security. While Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) have achieved high accuracy in this domain, CNNs often require millions of parameters and substantial computation. ViTs suffer from the quadratic time and space complexity of self-attention (SA), limiting their use on resource-constrained devices. Although SA is capable of modelling long-range dependencies when symptoms are dispersed, many plant diseases exhibit small, localized lesions or texture changes; therefore, Neighborhood Attention (NA) offers a more efficient and targeted alternative by focusing on nearby regions rather than the entire image.

This work proposes a custom Localized NA block implemented in TensorFlow/Keras that operates directly on CNN feature maps, bypassing patch embedding and transformer modules. A lightweight CNN is then developed by combining depth-wise separable convolutions with the proposed localized NA block. In addition, a 100-category plant disease dataset covering 16 crops is presented. The dataset is curated, class-balanced, and made publicly available to support reproducibility and encourage further research.

The proposed 9-layer CNN, with just 1.7M parameters and a size of 6.74 MB, achieved a favorable balance between accuracy, model size, and computational efficiency, compared with MobileNetV1, MobileNetV2, DenseNet121, InceptionV3, MobileViT-XXS, and EfficientViT-M0, achieving 98.97%± 0.33% accuracy on PlantVillage and 93.36%± 0.28% on the proposed dataset. The ablation study showed that the NA block improved test accuracy by approximately 2–3%, while Grad-CAM visualizations indicated more precise targeting of diseased areas in the leaf image.

This paper presents a set of feature selection methods for intelligent software performance monitoring based on machine learning models, with a focus on improving interpretability, scalability, and adaptability in high-dimensional telemetry analysis. The research addresses limitations of traditional statistical and rule-based approaches, which are often unable to capture nonlinear dependencies and dynamic interactions in modern distributed architectures. A unified methodology is proposed that integrates several complementary techniques for adaptive feature selection in intelligent monitoring systems. These include a topology-aware method based on graph neural networks for modeling structural dependencies in microservice architectures, a correlation-driven approach for reducing feature redundancy, a multifactor fusion method combining statistical significance, temporal stability, and predictive contribution, a cost-efficient strategy for serverless environments, and a context-aware reinforcement learning approach for dynamic feature adaptation. The proposed methods are evaluated on a large-scale dataset exceeding 3.5 TB, collected from 42 real-world applications representing monolithic, microservice, cloud-native, and serverless architectures. The results show an average reduction in feature dimensionality of 37%, while maintaining over 95% predictive accuracy across multiple models. Additional improvements include, on average, a 21% increase in dependency modeling accuracy, an 18% gain in feature relevance estimation, a 26% reduction in feature instability under dynamic workloads, and up to 42% cost reduction in serverless environments, as observed across repeated experiments under controlled workload variability and consistent evaluation settings. While the results demonstrate the effectiveness of adaptive feature selection, further validation in diverse real-world conditions is required to confirm the generalizability of the proposed framework.

Potholes are a major concern for road infrastructure, traffic safety, and vehicle maintenance. Manual inspection methods for pothole detection are labor-intensive, time-consuming, and often inefficient for large road networks. This study evaluates and compares the performance of YOLOv5 and Single Shot Detector (SSD) models for automated pothole detection under diverse weather and lighting conditions. Using the Multi-Weather-Based Dataset (MWBD), images captured during daytime, twilight, and nighttime were annotated with bounding boxes and enhanced through data augmentation techniques such as shearing and flipping. Experimental results indicate that YOLOv5 achieves a precision of 92.2%, recall of 89.2%, F1-score of 90.7%, and mAP@0.5 of 90.0%, while SSD achieves a precision of 88.5%, recall of 92.0%, F1-score of 90.2%, and mAP@0.5 of 91.4%. The comparative analysis demonstrates that both models are effective in detecting potholes across varied road textures and environmental conditions, with trade-offs between precision and recall. This study highlights the suitability of deep learning-based object detection models for automated road inspection, reducing human effort, enhancing maintenance efficiency, and improving road safety. The novelty lies in the systematic comparison of YOLOv5 and SSD under multi-weather conditions, providing practical guidance for intelligent transportation systems.

Hand-drawn circuit diagrams must be manually converted into hardware description languages (HDLs) for digital design workflows. This manual conversion is time-consuming and error-prone and there has been little focus on hardware validation along the entire end-to-end circuit design process (such as circuit recognition and code generation). In response to these challenges, we present Sketic-FPGA, an end-to-end machine learning-based automated framework for converting hand-drawn logic circuits into functionally verified implementations on FPGA devices. The Sketic-FPGA system operates in a six-stage pipeline consisting of: adaptive image preprocessing, gate detection using an improved Faster R-CNN with ResNet-50 backbone, topology extraction, synthesis-aware Verilog code generation, automated FPGA implementation using Xilinx Vivado toolchain, and hardware-level validation. The proposed model was trained with 800 annotated samples across eight classes of logic gates, utilizing rotation-aware detection and curriculum learning to improve robustness. When evaluated against a dataset of 200 previously-unseen, test circuits, Sketic-FPGA produced 99.2% detection accuracy and 98.8% classification accuracy. All designs generated with Sketic-FPGA were successfully synthesized and implemented onto actual FPGA hardware, achieving functional correctness across the entire test circuit dataset using LED testing, Integrated Logic Analyzer (ILA) waveform verification, and exhaustive truth-table validation. On average, processing each circuit took 29.4 seconds from start to finish which has reduced the time required for a designer to create a circuit manually. An examination of how long it took students to design a circuit revealed that they spent 67% less time across multiple design iterations. Although we have only demonstrated the effectiveness of our framework on combinational circuits and in a controlled environment, our results indicate that there are many opportunities for rapid prototyping and automated hardware design as well as support for digital educational methods.

Standard image denoising algorithms often rely on data-loss techniques. While fast, they perform poorly in low signal-to-noise ratio scenarios like astrophotography. Stacking-based algorithms resolve this by utilizing noise variability across multiple frames, but they require precise image alignment to prevent artifacts.

In this paper, we propose M-BRISK, a modified BRISK feature detection and description algorithm offering improved robustness and accuracy. A novel keypoint-detection strategy with efficient filtering contributes to this enhanced performance. Crucially, M-BRISK targets uniform spatial distribution of both keypoints and matched descriptor pairs across the image, achieved by partitioning the image into regions and enforcing balanced selection within each. This ensures that the estimated homography is informed by correspondences spread across the entire image rather than clustered in salient areas, leading to more reliable homography estimation using RANSAC.

Evaluation across four test images and four synthetic homographies simulating realistic camera displacements, as well as rotation invariance tests from 0° to 180°, demonstrate that M-BRISK achieves homography estimation accuracy superior to BRISK by 15.7% and to SIFT by 13.1%, while also exhibiting better average rotation invariance. Furthermore, M-BRISK maintains stable processing speeds regardless of keypoint count. While BRISK's performance degrades as detections increase, M-BRISK becomes substantially faster beyond 15,000 keypoints. As a result, M-BRISK enables fast, high-quality image registration, making it well-suited for subsequent denoising through stacking.

Accurate identification of insect pests is crucial for effective agricultural management and prevention of crop losses. Among these pests, tephritid fruit flies significantly impact fruit and vegetable production, leading to economic losses and reduced market quality. Existing insect identification methods largely rely on manual inspection by taxonomists, which is time-consuming, error-prone, and not feasible for real-time applications in field conditions. Moreover, many existing machine learning-based approaches suffer from limited generalizability and dependence on controlled environments, restricting their practical deployment. To address these challenges, this study proposes a Modified Convolutional Neural Network (MCNN)-based approach for automated identification and classification of tephritid fruit fly species. The proposed method integrates image segmentation, feature extraction, and data augmentation techniques to enhance classification performance under varying conditions. A real-world dataset was collected using pheromone traps from multiple agricultural locations in Punjab, India, comprising four major species: Bactrocera dorsalis, Bactrocera zonata, Zeugodacus cucurbitae, and Zeugodacus tau. The MCNN model is trained using optimized hyperparameters, including learning rate, batch size, and optimizer selection, to improve robustness and accuracy. Experimental results demonstrate that the proposed model achieves an accuracy of 90%, along with improved classification capability compared to traditional approaches. The integration of real-field data and enhanced preprocessing techniques makes the proposed system suitable for practical deployment in precision agriculture. This study contributes to the development of an efficient, scalable, and automated insect identification framework that can assist farmers and agricultural experts in timely pest management.

Chest digital tomosynthesis (DTS) provides a compromise between conventional radiography and computed tomography in terms of radiation dose and diagnostic capability. Most reconstruction algorithms used in DTS assume a stationary object during acquisition. However, projection data are acquired over a finite time interval, during which internal anatomical structures may exhibit temporal variability. In this study, projection-domain intensity variations were analyzed to assess temporal consistency of DTS data. Mean intensity values were measured across multiple regions of interest (ROIs), forming temporal intensity profiles for anatomically distinct regions. Additionally, intensity profiles across anatomical transitions were evaluated in both projection data and reconstructed slices. The results show that while global intensity variations are primarily driven by acquisition geometry, certain regions exhibit local fluctuations, indicating reduced temporal consistency. Comparative analysis revealed that regions with increased variability correspond to degraded contrast and broadened transition boundaries in reconstructed slices. In particular, the heart–lung interface showed a significant contrast reduction compared to the stomach–lung interface, despite similar contrast levels in projection images. These findings indicate that even small temporal inconsistencies in projection data can lead to cumulative reconstruction errors. The proposed ROI-based analysis provides a simple approach for identifying such regions directly from projection data and suggests directions for improving reconstruction quality, including more consistent 3-D reconstruction of cardiac slices across different phases of the cardiac cycle.

Rice isn't just essential to global food security; disease outbreaks and poor soil health management stymie its growth. Often, traditional rice farmers lack the knowledge or resources to diagnose plant diseases and monitor soil conditions in real time. Existing solutions typically focus on either soil health monitoring or disease detection, but not both, leaving farmers unable to respond to the identified threat. This project will mitigate this by combining real-time monitoring, disease classification, and a recommendation system into a single solution. Poor disease detection and inadequate health examinations of soils commonly result in decreased rice productivity. The proposed research is also focused on the development of an intelligent system equipped with IoT sensors to monitor soil parameters such as moisture, nitrogen concentration (NPK), and temperature in real time, and a machine-learning-based system capable of classifying 15 different rice diseases. The system also includes a recommendation engine that provides actionable recommendations for treating an illness, making it a complete soil and crop health management tool. The system is based on a transfer learning model (MobileNetV2) that classifies rice illnesses using image classification. The model was trained on 22,688 images of rice diseases, achieving a detection accuracy of 96.42%. The system was also highly accurate for monitoring soil health, with minimum standard deviations of 0.20% and 0.22 for soil moisture and nitrogen levels, respectively. The results obtained reflect the effectiveness of the developed system in enhancing the farming process by enabling farmers to identify diseases at early stages and improve soil conditions. Lastly, the methodology enhances rice production, reduces crop losses, and helps achieve global food security.

Video clip extraction is the process of generating shorter, focused video segments by identifying and retaining frames that contain a user-specified object of interest (UOoI). Such targeted extraction allows users to access relevant portions of a video without watching the entire recording, with practical use in surveillance review, content management, and educational settings. In this work, we present an object-conditioned video clip extraction framework that uses the pretrained YOLOv7 detector to perform frame-level analysis of an input video. For each frame, the detector produces a set of class labels, which are matched against the user-selected UOoI to produce a binary per-frame detection signal. A one-dimensional temporal-window voting filter is then applied to this signal to suppress isolated false positives and recover isolated false negatives, addressing the single-frame detection noise that produces visible discontinuities in naive frame-by-frame approaches. The voted-positive frame indices are mapped back to source timestamps, and the corresponding audio segments are extracted directly from the source video using ffmpeg, preserving the original audio track in the output clip. The framework uses a dictionary of 80 object categories drawn from the MS COCO label set, and a graphical user interface allows users to select an input video, choose a target object, preview the input, and obtain the extracted clip with audio. We evaluate the framework on the SumMe benchmark, which we re-annotated at the frame level for object presence, and on a newly annotated set of 39 videos collected from public sources. Both datasets were independently labelled by two annotators, with Cohen’s kappa of 0.85 and 0.83, respectively, and disagreements resolved by a third adjudicator. At the default voting configuration of W=5, K=3, the framework attains an F1-score of 70.88% with 90.12% accuracy on SumMe and an F1-score of 69.89% with 85.13% accuracy on the custom dataset. An ablation over voting parameters shows monotonic gains on SumMe across the full sweep, and a smaller, dataset-dependent gain on the custom set. We discuss the remaining limitations of the pipeline, including single-UOoI conditioning, dependence on the MS COCO label vocabulary, and abrupt transitions between non-adjacent extracted segments, and outline directions for addressing them. abstract.

Social media streams reflect opinion of public in real-time, but short and noisy tweets make it hard to attribute sentiment to entities and this paper introduces an AI/ML pipeline to classify the sentiment towards the referenced entities in twitter messages. Using the Twitter Entity Sentiment Analysis benchmark (twitter_training.csv and twitter_validation.csv), the tweets are normalized (lowercasing, punctuation, platform specific artifacts, tokenization, stop word filtering and lemmatization) and represented using TF-IDF (Term Frequency–Inverse Document Frequency) features with a maximum of 5000 terms. Machine learning models including Logistic Regression, linear SVM, Multinomial Naive Bayes, and ensemble and neural methods such as Random Forest, XGBoost, and Multilayer Perceptron (MLP) are trained on the training split and evaluated on the validation split using macro-averaged precision, recall, F1-score, and confusion matrix analysis. The results show that linear discriminative models are well suited to sparse TF-IDF spaces, with SVM and Logistic Regression providing balanced class-wise behaviour and Naive Bayes offering a computationally efficient baseline. XGBoost delivers moderate improvements over simple probabilistic models, while Random Forest achieves substantial gains through ensemble learning. The best overall performance is obtained by MLP, demonstrating that non-linear neural modeling more effectively captures complex feature interactions and entity-relevance patterns. Misclassifications are focused on the Neutral - Irrelevant boundary, resulting in instances of relevance ambiguity at the entity level in the case of sentiment and driving future extensions with context aware deep architectures and entity conditioned representations. These baselines provide support for monitoring purposes for brands and public figures as well as expose the limits of non-contextual features for sarcasm and implicit targets in Twitter discourse.

As EV fast charging infrastructure has developed rapidly, electromagnetic compatibility (EMC) problems with DC-DC high-frequency converters have become more acute. In particular, multi-source multi-level active bridge (MS-ML-DAB) converters operating over a wide range of output voltages experience significant electromagnetic interference (EMI) due to rapid switching transitions and parasitic components. This study systematically analysed the electromagnetic interference (EMI) generated by the common mode (CM) and differential mode (DM) of the MS-ML-DAB converter for EV DC fast charging systems. A parasitic converter model is proposed to precisely identify the major noise processes in the EMI. Using FCC Class B emission-guided and impedance-based design concepts, the multi-stage EMI filter of the type π is analytically constructed for both input and output ports. The recommended filter shall provide impedance matching, adequate damping and stability in the event of significant voltage and current fluctuations. PSIM based simulations show significant noise abatement of DM and CM in the frequency range 150kHz to 30MHz in both working modes, confirming compliance with FCC Class B standards. This technique offers a practical and scalable solution to mitigate the EMF in high-power EV chargers.

This work aims to preserve and promote the rich musical heritage of Assam by developing an automated classification system for Assamese regional songs using a hybrid deep learning approach. This method not only modernizes the preservation of traditional music but also enhances its accessibility to a global audience for integrating AI with cultural conservation. Five genres of Assamese songs—Bihu, Kamrupiya Lokageet, Goalporiya Lokageet, Borgeet, and Naam—are considered in this study. By leveraging Convolutional Neural Networks (CNNs) and advanced audio feature extraction techniques such as Mel-Frequency Cepstral Coefficients (MFCCs) and spectrograms, a hybrid model combining VGG16 and ResNet50 is developed. This fusion utilizes the strengths of both architectures, enhancing the model’s performance and accuracy. Following the process, it is observed that two distinctly different genres, Bihu and Borgeet, are accurately categorized by the proposed model, while the remaining three show slight labeling inconsistencies.

This paper presents a comparative study for disease detection that proposes to combine machine learning classifiers (kernel support vector machine, random forest, decision tree, and eXtreme gradient boosting) to form a stronger ensemble classifier and also deep learning classifiers (long short-term memory and convolutional neural networks) to make a decision on whether the deep learning classifiers individually work better or the ensemble classifier consisting of four machine learning classifiers following the feature extraction method bag of features. The main reason for the global crisis in agricultural production is the presence of various vegetable diseases. It damages food quality and reduces production. These diseases must be detected, which is a challenging task to perform manually. Using various algorithms, we can identify vegetable diseases. Recently, deep learning has demonstrated notable success in the field of precision agriculture for identifying vegetable diseases. In this paper, the detection of vegetable diseases is done using three techniques: Ensemble Machine Learning Classifier (Kernel support vector machine, Random Forest, Decision Tree, and eXtreme gradient Boosting), CNN, and LSTM. By using the Bag of Features feature extractor, we extract 500 feature words from each vegetable dataset. CNN, LSTM, these two deep learning algorithms, and the ensemble method of machine learning classifier are used to make classifications of healthy and disease-affected vegetable categories and generate a confusion matrix. Then, from the confusion matrix, the performance metrics (precision, recall, F1-score, and accuracy) are identified. By applying soft voting for each individual classifier of machine learning, we predict the average best accuracy for each of the datasets. At the end, compare the performance of the ensemble method with the two deep learning algorithms according to the accuracy value. For the Cauliflower dataset, the Ensemble Machine Learning Classifier gives the accuracy of 83%, the deep learning classification algorithm CNN presents the accuracy of 94.51%, and LSTM gives the accuracy of 92.95%. The potato dataset's ensemble method accuracy is 89%, Convolutional Neural Network's is 89.47%, and Long Short-Term Memory's is 84.35%.