Context. Automated wood species recognition from macroscopic sections must combine high accuracy with stringent computational and memory requirements typical of peripheral devices and UAV onboard platforms. The heterogeneous, partially self-similar structure of wood texture, combining self-similar mid-level fluctuations with locally high-contrast elements and a large spread of leaders reflecting multifractality, motivates the creation of a fractal neural network architecture that can encode multiscale patterns while remaining computationally efficient.Objective. The goal is to develop and validate fractal neural network models that achieve high recognition quality at reduced computational cost, enabling practical deployment on systems with limited resources.Method. A fractal neural architecture (FractalNet) is employed to realize a self similar multi branch topology that forms an ensemble of receptive fields from fine to coarse scales, coherently modeling the spectrum of local regularities captured by leaders. The approach is benchmarked against ResNet50 and VGG16 with and without data augmentation on a 12 class macroscopic image task. Evaluation includes per class precision, recall, and F1, macro/weighted aggregates, confusion matrices, and analysis of numerical complexity in terms of trainable parameters and layer depth to assess deployability.Results. The FractalNet combined with augmentation procedure attains the best overall performance, reaching macro F1 = 0.80, weighted F1 = 0.81, and accuracy = 0.81, outperforming ResNet50 (macro F1 = 0.57) and VGG16 (macro F1 = 0.71). Confusion matrices exhibit reduced cross class confusions, indicating more uniform gains across species. Despite superior accuracy, FractalNet contains ~0.37 M parameters versus 23.6 M in ResNet50 and 65.1 M in VGG16, yielding a markedly smaller memory footprint and lower inference latency.Conclusions. The alignment between multifractal texture properties of wood and the self similar design of FractalNet produces a favorable accuracy-efficiency trade off. The method delivers state of the art recognition quality while preserving computational frugality, thus enabling reliable use in resource constrained scenarios, including UAV based and other edge deployments
Context. Optimizing search in multidimensional catalogs of radio-electronic components (sensors, microcontrollers, communication modules) is a crucial task for computer-aided design systems, electronic component logistics management, and intelligent technical support systems. The complexity arises due to the high dimensionality of the parameter space (operating frequencies, power consumption, temperature ranges, etc.), data heterogeneity, and the high frequency of complex queries combining numerical constraints with categorical filters. Classic indexing algorithms from relational database management systems are inefficient for this specific domain, which slows down the performance of real-time information systems. Modern indexing technologies for searching multi-attribute catalogs of technical components are moving away from the paradigm of universal one-dimensional structures in favor of specialized and hybrid approaches tailored to the nature of technical data. The focus has shifted from fast single-key search to efficient pruning of multidimensional parameter space. To achieve this, spatial indexes are actively used, interpreting each component as an object in an N-dimensional space, where each technical parameter is a separate axis. This allows a single query to the index to find all entries that fall within a specified multidimensional hyperrectangle. Simultaneously, technologies that treat search as an information retrieval problem are evolving. Categorical attributes, such as interface type or manufacturer, are indexed using inverted indexes or compressed bitmap indexes, which provide ultra-fast execution of AND/OR operations over large sets. A separate direction involves the use of vector representations (embeddings) of technical characteristics, obtained using machine learning models, followed by indexing using specialized structures for nearest neighbor search. This enables semantic search by technical description or finding analogous components. A key issue in indexing for search within multi-attribute catalogs of technical components is the selection and combination of data structures that effectively prune the search space across all relevant dimensions simultaneously, minimizing the retrieval of iirrelevant data in the early stages of query execution.Therefore, the development of a new method that systematically combines the strengths of modern approaches within a unified adaptive architecture is a relevant scientific and technical problem. Its solution will significantly reduce the execution time of complex queries in key information systems for the fields of radio electronics, telecommunications, and automated engineering, addressing the challenges of production digitalization and intelligent data processing.Objective. Development of a combined indexing method for efficient execution of complex search queries in multi-dimensional catalogs of technical components.Method. A combined indexing method is proposed, which combines R-tree for multidimensional filtering of numericalparameters and inverted indexes for categorical features. The method for improving search efficiency is based on an adaptive query planner that dynamically chooses the optimal execution strategy (Index-First, Parallel-Merge or Full-Scan) based on an assessment of the selectivity of the conditions.Results. The problem was formulated, and a combined indexing method was developed for multidimensional catalogs of radioelectronic components. In the course of the study, a formalized mathematical search model was created that takes into account the selectivity of numerical and categorical conditions, and algorithms for automatic distribution of attributes between index types and dynamic selection of the query execution strategy were developed. A hybrid index architecture was proposed that combines an R-tree for indexing numerical parameters and inverted indexes for categorical features, as well as mathematical models for assessing selectivity and optimizing the query execution plan. Experimental studies on a synthetic data set confirmed the effectiveness of the developed method, demonstrating a reduction in the execution time of complex queries compared to basic indexing based on B-trees.Conclusions. The work develops a combined indexing method for efficient execution of complex multi-attribute queries in radioelectronic component catalogs. The method is based on a formalized mathematical search model, which allowed building a hybrid index architecture. This architecture integrates an R-tree for filtering numerical parameters, inverted indexes for categorical features, and an adaptive planner that dynamically selects the optimal query execution strategy (Index-First, Parallel-Merge, FullScan) based on an assessment of the selectivity of conditions. An algorithm for the automatic distribution of attributes between index types is developed. Experimental modeling confirmed the effectiveness of the method, showing a 35–55% reduction in the execution time of complex queries compared to traditional B-tree-based indexing, especially for queries typical of engineering component selection. The results obtained prove the feasibility of using combined indexing to increase the productivity of information systems working with multidimensional technical catalogs
Context. The relevance of the article is due to the need for further development of simple algorithms for analyzing the structural connectivity of network objects, in reducing the computational complexity and increasing the functional capabilities of such algorithms. The linear algorithm proposed in the article can be applied: in urban and rural transport route planning systems; in solving various types of problems related to carrying out restoration work on network objects after their destruction and the loss of certain connections between the elements of such objects (power supply systems, communication systems, transport networks, etc.); in developing options for effectively destroying key objects of the enemy’s critical infrastructure; in analyzing social networks for grouping individuals into groups by interest, etc.Objective. The goal of the work is to develop a linear algorithm for searching for strongly connected components in the structure of network objects, which is resistant to changes in the structure of objects and has a polynomial computational complexity.Method. The physical object to be studied is modeled by a connected directed graph. To determine the structure of the model graph components of strong connectivity, the idea of analyzing the presence of direct and reverse transitive closures between all pairs of vertices of the model graph was applied. Since there is a route (or arc) connecting them between all pairs of vertices included in any component of strong connectivity, then in the matrix of transitive closures of binary relations of such a graph there will always be unit blocks (transitivity compactions). Individual elements of these blocks in the matrix of transitive closures are scattered in a certain way. By analyzing the matrix for the presence of identical rows and identical columns, it is possible to establish their belonging to certain unit blocks. After such analysis and grouping the corresponding rows and columns into unit blocks, it is possible to determine the number of components of strong connectivity, as well as the index (vertex) composition of each component.To construct a matrix of transitive closures of binary relations, any of the known algorithms can be used, as discussed in the article. Therefore, the linear algorithm proposed in this article is conventionally divided into two main stages: the first is the search for transitive closures of binary relations of a model directed graph; the second is the determination of the matrix of inverse reachability of the graph, the determination of the matrix of mutual reachability, and the search for unit blocks in the latter by lining them up on the main diagonal of the matrix.Results. 1) A theoretical basis for analyzing the matrix of transitive closures of binary relations of a model directed graph has been developed in order to determine the quantitative and index composition of strongly connected components in its composition; 2) In terms of graph theory, the search for strongly connected components in the structure of a network object is formalized; 3) The developed algorithm was verified for its ability to determine strongly connected components in the structure of a directed graph.Conclusions. Theoretical studies and a number of experiments confirm the possibility of using the proposed algorithm in the tasks of structural analysis of network objects. Since the well-known provisions of graph theory were used in the theoretical studies, which are interpreted unambiguously in this theory, the absence of any probabilistic processes allows us to consider the developed algorithm accurate. Polynomial estimates of its computational complexity allow it to be used in real-time scale
Context. Ensuring robust, adaptable, and compute-efficient object detection in UAV aerial imagery under distribution shifts, structured and unstructured noise, and strict onboard latency/energy budgets is an urgent scientific task. A compute-aware detector and a complementary training/adaptation method that integrate a dynamic transformer backbone with gate units, parameter-efficient adapters, and resource-bounded test-time adaptation to sustain accuracy under realistic perturbations and domain shift.Objective. Development of a model and method for object detection in aerial imagery that jointly provide robustness and adaptability while meeting embedded compute and real-time constraints typical of onboard UAV systems.Methods. The approach combines dynamic neural networks with Gumbel-Softmax gate units over a ViT-T/16 backbone, a Simple FPN and a RetinaNet-like one-stage head, budget-aware losses that target a desired dynamic compression rate, structured procedural noise (Perlin, Gabor, Worley) for robustness training, LeakyReLU6 with a straight-through estimator for stable gradients, and test-time adaptation via objectness-weighted marginal-entropy minimization on lightweight adapters.Results. On VEDAI, a gated ViT-T/16 detector reaches mAP@0.5 of 0.77 at ~5.0 GFLOPs and 17.8 FPS, rising to 0.79 with adapters and TTA, whereas a static counterpart attains 0.74 at 9.6 GFLOPs and 10.6 FPS; pretraining with procedural noise lifts accuracy further to 0.80 (gated) and 0.82 (gated+TTA) with minimal compute overhead. Under domain shift (trained on VisDrone, evaluated on VEDAI), dynamic gating and TTA improve mAP from 0.54 to 0.60 without noise pretraining and up to 0.66 with it, sustaining ~5.4–5.6 GFLOPs and ~16–17 FPS within an 8–10 GFLOPs budget on 4×A76 CPUs.Conclusions. The proposed object detection model and method – combining dynamic gating, perturbation-aware training, andbudgeted test-time adaptation – reduce average compute while increasing robustness and adaptability, yielding a superior accuracythroughput trade-off for UAV onboard deployment under real-world disturbances and distribution shifts.
Context. The modern information society, with its rapidly expanding means of information exchange and globalized data, offers the opportunity to study information dissemination, investigate its impact, and create information security systems. Mathematical modeling methods are widely used in the development of information dissemination systems and technologies. Thermal, diffusion, and mechanical processes – that is, processes occurring in a continuous medium – are considered fundamental. Among the various modeling approaches, equations of mathematical physics, which formalize the fundamental laws of substance transfer, are often used.Objective. The aim of this paper is to develop a comprehensive, formalized approach to numerically modeling the dynamics of information dissemination processes in social networks based on the principles of cellular automata and diffusion models. The subject of the study is the analysis of the dynamics of the level of information generated based on the internal behavior of each network cell, assuming that its initial state is formed as a result of the influence of certain processes outside the information communities.Method. This paper proposes a hybrid approach to modeling information dissemination processes based on cellular automata, which observes the internal and external dynamics of individual cells. The state change of the automata model is described by a specified transition function and rules for generating output signals, while a diffusion approach based on heat conduction principles is used to formalize internal changes in the cell’s state. Computational experiments were conducted to model the dynamics of information dissemination processes, taking into account various types of external influences from others, and calculations of information dissemination dynamics indicators in social network communities are presented.Results. The proposed hybrid model allows us to describe and analyze information dissemination processes within a social group, which is formed, for example, from social network subscribers. The group is divided into subgroups that are relatively homogeneous in terms of specific indicators in the form of online communities.Information processes were modeled taking into account various external influences from surrounding environments (using von Neumann and Moore neighborhoods as examples) with explicitly defined transition and exit functions. Internal information dissemination processes in specific cells were formalized using scalar heat equations. Numerical calculations of information dissemination dynamics in each community and the group as a whole were obtained, and a solution to information content problems (participants’ attitudes toward a specific problem) in subgroups was proposed based on determining the locations of centers of iinfluence.Conclusions. This article examines the application of cellular automata principles to modeling the dynamics of information dissemination processes. A new approach is proposed that, in addition to the automata model, considers the internal behavior of each cell, assuming that its initial state is formed as a result of certain intracellular processes at each time interval. In other words, a hybrid version of a cellular automaton for observing the internal and external dynamics of individual cells is considered, whereby the change in the states of the automaton model is described by a given transition function with certain rules for generating output signals, and a “mechanistic” approach based on the principles of thermal conductivity is used to formalize internal changes in the state of the cell.
Context. Modern data mining methods are widely used to build classification and predictive models. However, when processing discrete data, the problem of quantitatively assessing the informativeness of features, which determines the accuracy and stability of classifiers, remains. The lack of a universal approach to measuring the contribution of both individual features and groups of features to the classification result complicates the process of automated feature selection and model optimization.Objective. The aim of this paper is to develop and theoretically substantiate a method for assessing the informativeness of both individual features and arbitrary groups of discrete features, based on the relationship between the statistical characteristics of features and measures of class distinguishability.Method. An approach is proposed that links the informativeness of both individual discrete features and arbitrary groups of such features with respect to a function characterizing the target variable. The method is based on the research results of renowned scientists Kendall and Stewart in the field of nonparametric statistics. For practical application, an algorithm for calculating the informativeness of both individual features and groups of features is introduced, suitable for implementation in data analysis software systems.Results. It is demonstrated that the developed method enables formal and quantitative evaluation of the contribution of both individual features and arbitrary groups of features to the classification process without prior assumptions about the model type. It also enables the identification of hidden dependencies between features, which is impossible with individual assessments; i.e., it enables the identification of feature interactions. The resulting expressions provide a basis for automating feature selection when working with discrete data, improve the analytical value of the method, and offer a basis for meaningful feature selection.Conclusions. The proposed approach enables the unification of the procedure for assessing the informativeness of bothindividual discrete features and arbitrary groups of features in data mining systems. It provides a formal link between the statistical characteristics of the data and the quality of the classification, which contributes to increased accuracy, robustness, and interpretability of models
Context. We investigate such orthogonal function systems as the Walsh function one and the Bessel function one for the searchof the weight function of the Kolmogorov-Wiener (KW) filter for the forecasting of the heavy-tail MFSD (multifractal fractional sum-difference) random process. The results for the MAPE (mean absolute percentage error) of the misalignment of both sides of the Wiener-Hopf integral equation are compared to that obtained by the Chebyshev polynomial expansion in our previous paper.Objective. The objective is to derive the weight function of the KW continuous filter via the truncated expansions in Walsh and Bessel functions and to compare results with the results obtained via the Chebyshev polynomials in our previous paper.Method. Galerkin method with the Walsh functions and the Bessel functions orthogonal on the required interval as the method basis is used.Results. It is shown that the choice of Walsh functions leads to better results for the above-described MAPE than the choice of the Chebyshev polynomials and Bessel functions. It is shown that the choice of the orthogonal polynomials is more effective than the choice of the Bessel functions. It is also obtained that of the approximation of 128 Walsh functions leads to the MAPE less than 0.5%.Conclusions. The weight function of the KW filter in the continuous case is investigated for the forecasting of a random stationary heavy-tail process in the MFSD model. The Walsh functions and Bessel functions are chosen to be the basis of the Galerkin method described in the paper. The results are compared to that obtained with the choice of the Chebyshev polynomials. It is obtained that the Walsh functions lead to the most reliable results among the above-mentioned orthogonal function systems. The calculated results may be applied for the practical forecasting of traffic in telecommunications and also they may be applied to the treatment of random processes in other fields of knowledge: in agriculture, etc.
Context. In recent decades, rapid advances in digital signal processing and artificial intelligence have greatly expanded capabilities in visual information analysis. A color image, as a complex multidimensional signal, carries geometric, spectral, and textural data about objects. Efficient processing of such data requires integrating filtering, segmentation, and transformation methods with machine learning algorithms to extract hidden patterns and meaningful features.Objective. The purpose of this work is to develop a method for analyzing color images based on quantization, binarization, clustering, and selection of priority clusters.Method. The proposed approach combines digital signal processing and machine learning to improve the accuracy and speed of extracting informative elements. It includes several interrelated techniques: Quantization of three-color components in the training template to reduce color diversity and accelerate binary image formation; selection of priority template colors based on occurrence probability with a normalized threshold, enhancing feature detection accuracy; quantization of the original image’s color components to optimize segmentation and avoid excessive clustering; construction of a binary image using quantized template colors to eliminate false clusters and improve clustering precision; extraction of binary elements via clustering, verifying only white-point surroundings to suppress noise and automatically identify elements of various shapes; selection of priority binary elements using probabilistic assessment to enhance reliability.Results. The method was implemented in Matlab and tested on a specialized database. Compared with traditional approaches, it demonstrated higher accuracy and stability in element extraction while reducing processing time through color optimization and removal of redundant clusters.Conclusions. The comprehensive application of quantization, binarization, clustering, and priority element selection ensures accurate, fast, and adaptive analysis of color images. The method expands the functionality of visual information processing systems and can be used for statistical analysis, intelligent image processing, segmentation, and classification of complex visual structures
Context. The study’s relevance stems from the urgent need to improve the effectiveness of prehospital first aid training under heightened risk to civilian populations, particularly in emergency scenarios involving damage to civilian infrastructure. Traditional training approaches are limited in their ability to realistically simulate hazardous situations, objectively monitor participants’ actions, and quantitatively analyse learning dynamics. Virtual reality (VR) technologies enable the creation of fully controlled and repeatable simulation environments with automated logging of temporal, behavioural, and performance-related parameters, providing new opportunities for objective assessment of training effectiveness.Objective of the study is to develop and experimentally validate an information technology for the quantitative evaluation of VR-based training effectiveness in developing prehospital first aid skills, compared with traditional training methods.Method. An experimental study was conducted using a controlled design with VR and control groups, including pre-test, posttest, and delayed retention measurements. Training effectiveness was evaluated using a set of quantitative metrics, including reaction time RT, action accuracy, number of critical errors, Precision, Recall, F1-score, and a composite performance score S. Learning dynamics were analysed using exponential learning curve models, mixed-effects models for repeated measurements, parametric and non-parametric statistical tests, bootstrap confidence intervals, and effect size estimation (Cohen’s d).Results demonstrate a statistically confirmed advantage of VR-based training over traditional methods. The average reaction time for critical actions in the VR group was reduced by approximately 10–20% compared to the control group (e.g., 34 seconds vs. 40 seconds in bleeding control scenarios). Action accuracy increased from approximately 0.78 in the control group to 0.86 in the VR group, corresponding to an improvement of about 8–10%. The composite performance score S was higher in the VR group by 0.05–0.12 (on a 0–1 scale), depending on the scenario. F1-scores for automated action classification reached 0.90–0.92, and large effect sizes were observed, with Cohen’s d values up to approximately 2.3. Retention testing further indicated improved stability and long-term preservation of skills following VR-based training.Conclusions. The proposed information technology and experimental results support the use of VR as an effective, scalable, and data-driven approach for prehospital first aid training for civilians, emergency responders, and medical personnel in emergency and disaster-response contexts.
Context. Semantic segmentation of teeth on panoramic X-ray images is an important task in dental diagnostics, as it allows for the automation of the diagnosis of dental diseases. However, panoramic X-rays have a complex structure, which complicates the task of segmentation. The use of convolutional neural networks shows high potential in solving this problem. In this context, it is relevant to study the best models and combine them into an ensemble in order to improve the quality of segmentation.Objective. The aim of this work is to study the effectiveness of various convolutional neural network architectures in the task ofsemantic tooth segmentation on panoramic X-ray images and to develop an ensemble approach to improve the quality of the results.Method. Various architectures of convolutional neural networks are used: U-Net, Attention U-Net, Residual U-Net, Residual Attention U-Net, R2 U-Net, U-Net++, U-Net 3+, USE-Net, Dense U-Net, and DeepLabV3+ with a pre-trained ResNet-101 backbone on the ImageNet dataset. An ensemble approach based on the best models is proposed, where the final segmentation mask is determined by majority voting. The models were trained on a preprocessed dataset of panoramic X-ray images with the application of augmentation techniques. The performance of the models was evaluated using the IoU, Dice, and Accuracy metrics.Results. Various neural network models were investigated, and the best ones were combined into an ensemble. The conducted experiments confirmed that the ensemble approach improves segmentation accuracy compared to individual models. The best result was achieved by the ensemble combining the Dense U-Net, Attention U-Net, and U-Net 3+ architectures.Conclusions. The proposed ensemble approach demonstrated high efficiency in the task of semantic tooth segmentation on panoramic X-ray images, outperforming the results of individual models. The scientific novelty of the study lies in the application of an ensemble approach that combines various architectures of convolutional neural networks for semantic tooth segmentation on X-ray images. The practical significance of the work is in the potential use of the developed approach for building automated diagnostic systems in dentistry. The obtained results can be applied to further automate the analysis of X-ray images and contribute to the development of intelligent medical systems.
Context. Proper and effective diagnosis of chest diseases is vital in timely treatment and efficient clinical decision-making. Chest X-rays (CXRs) are commonly utilized in the detection of chest diseases because they are readily accessible and cost-effective. Nevertheless, radiograph interpretation is still a time-consuming, subjective, and error-prone process, especially in health care settings where resources are limited. Radiologists require automated systems capable of producing consistent diagnostic information that will facilitate quicker and standardized patient care.Objective. The proposed research will develop and assess a deep learning model with the ability to produce valid and explainable diagnostic reports using chest radiographs. The main aim is to minimize human error, save time in the diagnostic process, and deliver uniform findings, which will guide clinicians to make sound decisions within a short period of time.Method. The images are then extracted using a convolutional neural network called GoogleNet that extracts high-level visual features, which contain important structural and anatomical information. The features extracted are then fed to a Long Short-Term Memory network, which represents the sequential character of the report generation process by conditioning itself on relationships between words and phrases in diagnostic text. To enhance its accuracy and interpretability, an attention mechanism is added to allow the system to concentrate on the most clinically valuable parts of the image when producing every part of the report. The Indiana University Chest X-ray dataset was employed to train and evaluate the proposed system, while several experiments were carried out to evaluate the performance of the proposed system regarding its performance against the existing benchmark models.Results. The GoogleNet-LSTM-Attention model was shown to be more effective at generating high-quality diagnostic reports. It has performed well compared to benchmark models on various natural language evaluation measures, such as BLEU, ROUGE, and CIDEr scores. These improvements indicate that the quality of clinical data and fluency of text generated are correlated and that CNN, RNN, and attention mechanisms are effective in medical image reporting.Conclusions. The study presented shows that a combination of CNNs, LSTMs, and attention in a single architecture has the potential to transform the process of interpreting chest X-rays. The system proposed not only improves the precision of the diagnosis but also provides clinical assistance, as it allows for performing radiographic assessment rapidly, consistently, and interpretably. Such AI-driven systems can bring about the potential to reduce workloads, decrease diagnostic errors, and enhance patient outcomes in various healthcare facilities.
Context. The research is devoted to solving the relevant scientific and applied problem of a background determining and identifying of the compatibility of HMI subjects based on irrational part of their interaction with HMI object(s). The object of research is a process of determining the compatibility of HMI subjects based on their perception subjectivization of HMI object. The subject of research are methods and means of artificial neural networks, computational intelligence, mathematical modeling, as well as computer programming. Objective. The goal of the work is the development of a HMI subjects’ compatibility determining method, based on their perception subjectivization of interaction object.Method. A method for determining the compatibility of subjects of HMI has been developed, which is based on the analysis of subject’s polyfactor portraits of perception subjectivization of the object(s) of this interaction, and ensures the possibility of solving the declared scientific and applied problem. The peculiarity of the developed method consists in the application (in the scope of solving the declared problem) of polyfactor portraits representing the perception’s subjectivization of the object(s) of HMI – by the subjects of the same interaction, in order to identify the degree/level of compatibility of these subjects specifically in the context of the prism of subjectivization features of their personalized perception of the object(s) (or processes) within the framework of their joint and common HMI. The proposed method is based on the appropriate models, developed and presented in scope of this work, as well as on a specialized algorithm that provides the possibility of automation and software/computer modeling of researched processes.Results. A basic mathematical model, a specialized algorithm, as well as a software model, have been developed and presented in scope of this research. Additionally, a practical approbation of the developed method has been carried out on the example of solving a relevant specific practical applied problem of identifying potentially interchangeable subjects of the researched HMI team, which confirms the effectiveness and efficiency of developed method in the context of solving relevant practical applied problems.Conclusions. A method for determining the compatibility of HMI subjects based on their polyfactor portraits of interaction objects’ perception subjectivization has been developed, which provides the possibility of solving the relevant scientific and applied problem of determining and identifying the compatibility of HMI subjects, and provides possibility(-ies) for identifying the compatibility of HMI subjects precisely on the basis of their polyfactor portraits of perception subjectivization of the object(s) of this interaction, thereby ensuring the improvement of the qualitative and efficiency indicators of HMI teams due to the possibility of identifying, selecting and joining subjects with a common subjective vision and perception of the object(s) and/or processes of their joint HMI.
Context. Numerical integration of rapidly oscillating functions of several variables is a key concept in engineering models and digital image processing. Despite the availability of various integration methods, the influence of smoothness and discretisation parameters on the accuracy of approximation remains insufficiently studied.Objective. The aim of this study is to analyze a cubature formula that uses economical interpolation schemes and to systematically investigate the influence of smoothness and discretisation parameters on the accuracy of numerical integration.Method. There are methods for numerical integration of rapidly oscillating functions of several variables, which are developedusing information operators that restore intermediate values of functions based on known values of the function at points, lines, and planes. Such information operators include the operators of O. M. Lytvyn, on the basis of which economical schemes for interpolating functions of several variables have been created. Their application in constructing cubature formulas for approximate calculation of double integrals of rapidly oscillating functions of several variables of general form allows calculations to be performed with high accuracy. The main focus is on the question of how the estimation of the error of numerical integration of two-dimensional rapidly oscillating functions in general form improves with the increase in the smoothness of the function.Results. The cubature formula of the approximate calculation of the double integral from the rapidly oscillating function of a general form is researched.Conclusions. A comparative analysis of the accuracy of the cubature formula for different classes of functions showed that the class of differentiability of a function is a determining factor that influences the rate of decrease of the theoretical error of numerical integration. Economical interpolation schemes and a higher level of smoothness of functions provide a significant increase in the accuracy of approximate calculation of integrals of two-dimensional rapidly oscillating functions of general form
Context. Finding solutions to problems with an incomplete set of necessary data is a relevant problem in various applied research. In object-oriented programming, a similar problem arises when it is necessary to create a class object in the absence of some data regarding its initialization.Objective. If there is enough data to solve a certain set of problems, it is advisable to create an object of the corresponding class. Problems arise when not all attributes of the created object have values. This can lead to errors in the operation of the object when calling methods that use undefined attributes. The aim of the work is to develop a mechanism that provides blocking of methods that are not ready to work, as well as their gradual unlocking as values for the attributes are defined.Method. A mathematical model of the class is proposed, which allows distinguishing two subsets of attributes that do not have values at the time of creating the class object. A method has been developed that automatically converts the source class, adding a mechanism for locking/unlocking certain methods depending on the presence or absence of attribute values that these methods directly or indirectly use.Result. The Class Transformer software product was created, which allowed a series of experiments to be conducted that confirmed the effectiveness of the research. The experiments showed a significant reduction in class conversion time compared to performing this work in the traditional way.Conclusions. The proposed method for automated class conversion in conditions of incomplete attribute definition, which, on the one hand, allows you to start working with the object without waiting for the moment of complete data definition, and on the other hand, reduces the time for class conversion. The method can be used for various object-oriented programming languages
Context. The increase in the number of wars in the modern world stimulates the progress of technological innovation, including the operation of drone systems and modern control systems. The research into mathematical simulations of pursuit, combined with methods of corporate differential games with dynamic constraints, can help throughout the development of the latest vehicle control systems in the civilian and military fields.Objective. The study is intended to develop a methodology for using multi-channel satellite imagery data in combination with the calculation of unpiloted aircraft motion and dynamic deformations of its elastic elements in vehicle guidance systems derived from pursuit models in cooperative differential game theory.Method. The methodological framework of the study consists in the integration of data from multi-channel aerospace images into the calculation of motion parameters and, based on this, the deformation of aircraft through the application of game theory methods and the development of advanced information systems to neutralize the opponent’s counteraction. The employment of aerospace photography has emerged as an essential element of many modern technologies in remote vehicle control systems. Despite the lack of reliable vehicle control systems using aerospace photography, they are frequently used and are instrumental in saving countless lives. Meanwhile, aerospace photography is exposed to a large number of excitatory factors that make information on them is commonly misleading, preventing its direct and correct application. The main strategy for increasing the reliability of processing and analytical results of aerospace images in this technique relies on multichannel images, i.e., multiple images of a single object acquired at different radiation frequencies, from distinct positions, angles, or time of shooting. This technique implements natureinspired strategies for optimal management, validated by wildlife evolution. The technology for processing multi-channel images and Integrating these results into vehicle management models remains underdeveloped and needs further refinement.Results. The study’s scientific novelty consists in establishing methodological foundations for utilizing data from multichannel aerospace imagery, computing unmanned aircraft dynamics within vehicle control systems, and developing advanced information technologies for optimal aircraft management. This approach is based on pursuit models in cooperative differential games and leverages neural network machine learning techniques.Conclusions. The experiments validate the effectiveness of the proposed approach for enhancing the accuracy of processing and analyzing aerospace images. A methodology for developing vehicle control systems based on pursuit models has been established. Future research will concentrate on adapting the model to new datasets.
Context. The rapid deployment of 5G networks and the emergence of 6G architectures introduce unprecedented traffic heterogeneity and burstiness across radio, edge, and core domains. Meanwhile, the energy footprint of mobile infrastructure is becoming a major sustainability concern, as carbon emissions increasingly shape network operation policies.Objective. This work aims to design a predictive carbon-aware multi-layer resource slicing framework for RAN – edge – core 5G/6G networks that jointly optimizes latency, cost, energy, and carbon emissions under bursty traffic conditions.Method. The proposed approach integrates an M/G/1-based queuing model for accurate representation of heavy-tailed service times and bursty arrival patterns; hybrid short-term/long-term forecasting of both traffic load and regional carbon intensity; and multi-objective optimization for carbon-aware VNF placement and traffic steering across network layers. A proactive – reactive orchestration mechanism performs predictive resource pre-allocation and runtime scaling.Results. Trace-driven simulations on a representative multi-layer testbed demonstrate a 34% reduction in CO2 emissions compared to latency-first orchestration, alongside a 22% decrease in operational cost and <1% SLA violation rate. Tail latency remains within slice-specific thresholds even under bursty loads, confirming that carbon reductions can be achieved without service degradation.Conclusions. Predictive, carbon-aware orchestration across RAN-edge-core domains substantially improves environmental and economic efficiency while preserving QoS guarantees. The results highlight the importance of integrating forecast-drivenoptimization and realistic traffic modeling into next-generation slicing architectures.
Context. Today, static difference colour models with integer coefficients are used to improve the efficiency of lossless imagecompression in graphic formats and archivers. These models improve compression, b ut do not consider the level of cross-correlation between different pairs of colour components of pixels in each image. Therefore, the development of methods for the formation and use of difference colour models for individual images in order to improve their compression by intercomponent decimation is currently an urgent scientific task.Objective. To develop methods and algorithms for the transition to difference colour models with integer coefficients and difference offsets to reduce compression ratios in the process of lossless compression of RGB images in modern graphic formats, in particular in PNG format.Method. Depending on the coding time constraints, the paper proposes to use 4, 16, 19 or 49 alternative difference colour models with difference offsets to select the most efficient model for each image. Prediction of the compression efficiency due to the use of the next alternative difference colour model is performed using entropy. The differences in the colour models are shifted so that the centre of the interval with the maximum number of these differences is shifted to the middle of the range of possible values. The effectiveness of three methods of determining the centre of this interval is investigated: without considering the deviations of component brightnesses, using the difference in component medians, and by determining the centre of the interval with the maximum number of component differences after their sequential search.Results. Our experiments have shown that, for example, applying difference colour models with integer coefficients to whole images in the process of sequential lossless compression, in particular, in the PNG graphic format we modified, allows reducingcompression ratios of photorealistic images of the ACT set by 0.19–1.06 bpb. Shifting the differences to the differences of the medians of individual components or centring the intervals of component differences provides an additional 0.01–0.02 bpb compression ratio reduction on average. Thus, difference colour models with integer coefficients and difference offsets can significantly increase the compression efficiency of lossless three-component photorealistic images in formats that use predictors and therefore can be implemented in the next versions of these formats at the standard level.Conclusions. In graphic formats, to reduce the lossless image compression ratio, in addition to decorrelation of individual component data, it is advisable to perform intercomponent decorrelation by switching to difference colour models with integer coefficients with difference offsets, which provide fast decoding. To maximise the reduction in compression ratio due to the application of the selected colour model, the midpoints of the difference intervals of the basic components R, G, B should be shifted to the middle of the range of possible values. When, for photorealistic images, due to strict limitations on encoding time or encoder size, it is impossible to select a difference colour model from among 49, 19 or 16 alternative ones, this choice should be made among three models: G – R + 128, G, G – B + 128; R, R – G + 128, B – G + 128, or R – G + 128, B – G + 128, B. For synthesised images, it is not advisable to switch to difference colour models with integer coefficients
Context. Recognizing and analyzing changes in transport infrastructure using satellite imagery is important for urban planning, emergency management, military monitoring, and post-war reconstruction. Traditional methods of assessing the condition of roads and bridges, based on ground surveys and expert interpretation of aerial photographs, are labor-intensive, time-consuming, and potentially dangerous for personnel, especially during natural disasters or armed conflicts. In this regard, it is important to develop automated methods for detecting bridge damage using machine learning from publicly available satellite imagery.Objective. This work aims to develop a neural network method for recognizing and pixel-localizing damage to transport infrastructure using Sentinel-2 multispectral images and OSM vector data, based on a modified Attention U-Net architecture with adaptive spatial-feature weighting and the Dice Loss function to produce a probabilistic damage mask.Method. A method for recognizing damage to transport infrastructure in Sentinel-2 multispectral images and OSM vector data using a convolutional neural network is proposed. The method is based on the improved Attention U-Net neural network architecture, in which vector data of the transport network (linear road and railway axes, bridge contours) serve as a priori structural information about the geometry and spatial location of infrastructure objects. This information is used by the decoder’s attention mechanism to limit the formation of output features to spatial areas corresponding to infrastructure objects. As a result, the loss function is calculated only within the spatial area specified by the binary mask of infrastructure objects. It reduces misclassifications in adjacent areas and increases the accuracy of boundary delineation for damaged bridge sections.Results. Experimental studies have confirmed the effectiveness of the proposed method for automated detection of bridge damage on multispectral satellite images. The comprehensive use of spectral-temporal information, combined with binary masks of transport infrastructure, has reduced false classifications in adjacent areas and increased the accuracy of localizing damaged sections. According to the quantitative assessment results, the mIoU was 78.6%, the F1-score was 0.81, and the overall classification accuracy exceeded 93%.Conclusions. The experiments confirmed the effectiveness of the proposed method for automated recognition of damage to road and railway bridges in satellite images. The improved Attention U-Net architecture, which combines spatial attention mechanisms with prior structural information about the transport network, has increased the accuracy of pixel-level damage area recognition compared to U-Net models. The proposed method enables the creation of analytical geospatial maps of damaged bridge sections, which can be directly integrated into geographic information systems for infrastructure monitoring, assessing the consequences of natural or military disasters, and supporting decision-making on response and recovery
Context. The problem of the parallel implementation of two algorithms for algebraic synthesis of a finite state machine with datapath of transitions is considered. This type of state machine can be used as an alternative to a finite state machine with a canonical structure in order to reduce hardware expenses in the state machine circuit. The object of the study is algorithms for finding complete and partial solutions to the problem of the algebraic synthesis of a finite state machine, which have a parallelimplementation based on a mechanism if processes. The first of these algorithms is the known algorithm for the complete sequential enumeration of state encoding variants with a fixed set of transition operations. The second algorithm implements an infinite enumeration of state encoding variants based on pseudo-random encoding. The goal of each algorithm is to find a solution to the algebraic synthesis problem with as few uncovered transitions as possible in a given time. This paper proposes an approach to increasing the speed of these algorithms, which consists in their parallel implementation using all processor cores available on the computer. This contributes to finding more efficient solutions to the algebraic synthesis problem in a given time, which can lead to lower hardware expenses in the device circuit.Objective. Multiprocess implementation and research of algorithms for finding solutions to the algebraic synthesis problem offinite state machine with datapath of transitions.Method. The research is based on the structure of a finite state machine with a datapath of transitions. The synthesis of the state machine circuit is preceded by a stage of algebraic synthesis, the result of which is the combination of a certain method of state encoding with the assignment of certain arithmetic-logical operations to individual transitions of the finite state machine. Such a combination is a solution to the problem of algebraic synthesis of a finite state machine with a datapath of transitions. In the general case, for a given finite state machine there are many solutions, each of which can be either complete (each transition is covered by one of the given operations) or partial (when part of the transitions remains uncovered by any of the operations). The more transitions in a partial solution are covered, the less hardware expenses are spent on implementing the state machine circuit and the better the solution found. The search for the best solutions requires an enumeration of a large number of possible variants of state encoding. In this case, it does not matter in principle whether the enumeration of the encoding variants is carried out sequentially or in a pseudorandom manner. In this work, in order to speed up the search for solutions to the algebraic synthesis problem, a parallel implementation of two algorithms is proposed, for which the “multiprocessing” module of the Python language is used. Both algorithms (the sequential search algorithm and the pseudo-random search algorithm) were implemented in programmatic way and investigated on the example of an abstract control algorithm using an i5-13500 processor. The purpose of the experiments was to evaluate the improvement of solutions to the algebraic synthesis problem found in the same running time by single-process and multi-process implementations of the specified algorithms.Results. Using the example of an abstract control algorithm, it is demonstrated that, in general, multiprocess implementation ofthe considered algebraic synthesis algorithms allows finding better solutions to the algebraic synthesis problem in the same time than with a single-process (non-parallel) implementation of these algorithms. The advantage of the parallel implementation of the algorithms is preserved when using different sets of transition operations.Conclusions. The algebraic synthesis of a finite state machine with a datapath of transitions is based on an algorithm for finding solutions to the algebraic synthesis problem. The paper proposes modified versions of previously known algorithms for finding such solutions, based on the use of the “multiprocessing” module of the Python language. The software implementation of these algorithms has proven that such an approach is generally better than a single-process search for state encoding variants, since it allows finding better solutions (solutions with fewer uncovered transitions) in the same time. The disadvantage of the proposed algorithms can be considered the use of more computer resources, which can negatively affect energy consumption
Context. Assessing text quality is essential for reliable AI that processes language. In ASR, it reflects how faithfully speech becomes text; in OCR, how accurately images yield text; and in NLP, how correct and coherent outputs are. Objective. The goal of the work is the creation of a complex metric for text quality assessment. Method. Classic metrics WER and CER are narrow: they capture only lexical edits, weigh all changes equally, ignore context and semantics, and often skip punctuation and case, masking readability issues and error types. We propose WELER, a hybrid metric that blends weighted WER and CER with a semantic component based on contextual embeddings to measure meaning preservation. Weights can be set manually or learned (e.g., via PCA), adapting the metric to ASR, OCR, or NLP tasks. Key challenges include computational cost, choosing optimal weights through correlation with human judgments, and the need for high-quality reference data. Proposed WELER metric integrates accurate word and character level error counting, using Levenshtein distance as a basis, with advanced semantic similarity methods based on contextual embeddings. This allows WELER to take into account not only what was incorrectly recognized, but also how much this error affects the meaning and understanding of the text. The inclusion of selfadjusting weights depending on the text category is a key feature of WELER, which allows adapting the metric to the specific requirements of different applications and domains, prioritizing those aspects of quality that are most critical for a particular task. Results. Proposed WELER metric is an alternative solution in this direction. It integrates accurate word and character level error counting, using Levenshtein distance as a basis, with advanced semantic similarity methods based on contextual embeddings. Conclusions. WELER, like all metrics based on reference data, relies on accurate and consistent human-verified transcriptions. Errors in the reference data can affect the accuracy of the assessment. Therefore, for complex metrics, the quality and representativeness of these data are especially important, since semantic and weighted errors are much more sensitive to the quality of the annotation than simple word counts.
