In the paper, existing methods of data augmentation have been analyzed and their shortcomings associated with limitations when operating with images obtained using unmanned aerial vehicles have been identified. A new specialized method for contextual augmentation of small objects, which supports spatial realism and is capable of providing the required density of small objects at the training stage of neural network models, has been proposed. Given the complexity of implementing the proposed method of contextual augmentation of small objects, it has been decided to divide the development process into several stages. To test empirically the proposed method, the Single Shot Detector MobileNet V2 model located in the TensorFlow 2 Detection Model Zoo repository has been selected as the base detector. This neural network model has been trained on the VisDrone dataset using standard data augmentation methods and the developed method. The obtained detection results of the basic neural network model have been compared with the improved model in accordance with the evaluation protocol of the Common Objects in Context Evaluation metrics library. Experimental studies have shown that context-aware augmentation significantly improves the ability of neural network models to detect small objects. The obtained result is of great practical importance when operating with compact energy-efficient systems with limited performance and memory.
Detecting plant diseases is a challenging problem for farmers, but computer vision and image processing are being offered as solutions. The main problem is the limited information that can be obtained from an image alone. Using computer vision alone does not take into account weather conditions and visual similarities between symptoms of various diseases. These problems can be solved by developing an expert system using an ontology containing knowledge about plant diseases, pathogens, and symptoms. We developed an ontology of plant diseases by integrating existing ontologies and adding disease-causing factors. Our Plants and their Diseases Ontology can improve diagnostic accuracy by including detailed descriptions of symptoms and their association with specific pathogens. Wikidata served as the primary source of taxonomic data, and queries against the data allowed for the extraction of relationships between plants, pathogens, and diseases. Using data from the American Phytopathological Society and the European and Mediterranean Plant Protection Organization, we were able to identify relationships not previously documented in Wikidata. Furthermore, the use of large language models helped to extract various symptoms of pathogens that cause plant diseases. The final ontology consists of 5002 classes and eight properties that relate various entities, including plants, pathogens, and symptoms.
This article continues the development of the Descriptive Theory of Image Analysis proposed by the authors and continues a series of articles on characterizing the image formalization space as a mathematical object—a mathematical space. The image formalization space is defined with the aim of formalizing and standardizing the description of all admissible states of image representations and procedures for bringing images to a recognizable form: a single model corresponding to a certain mathematical space and allowing its implementation in therein. The main result is the definition of the image formalization space as a mathematical space and the hierarchical structure of its organization.
The problems related to the assessment of pronunciation accuracy in speech, identification and classification of linguistic accents, and identification of the speaker’s language have been intensively studied in recent decades. The development of neural network technologies and the emergence of open source large language models have created new opportunities for the successful development of this field. In this paper, the methods for assessing the pronunciation quality of Russian words by Chinese speakers using information from a large, self-supervised multilingual model are considered. Algorithmic assessments of pronunciation quality are proposed and their correlation with expert (human) assessments is evaluated. It is shown that algorithmic assessments of pronunciation quality can be considered as more rigorous and competitive. Using a dataset of speech recordings of Chinese and Russian speakers, experimental values for indicators of Russian word pronunciation quality among Chinese speakers were obtained. The results show, for example, that a significant percentage of regular pronunciation errors can be interpreted as the devoicing of voiced consonants. These data are consistent with expert knowledge. The algorithmic criteria obtained can be useful for the development of pronunciation training methods and software, particularly for providing learners with effective feedback, conveying information about the pronunciation quality of individual words and localizing pronunciation errors.
In this paper, an approach to automated classification of brain magnetic resonance imaging images according to Alzheimer’s disease stages using explainable machine learning methods has been proposed. A comparative analysis of neural network architectures, which resulted in the development of an effective model that achieves 95
The Visual Genome dataset is currently among the select few that employ scene graphs to describe images, i.e., in addition to providing information regarding the objects present in the image and their attributes, it also elucidates the relationships between these objects. This makes Visual Genome a promising foundation for new datasets, including those in other languages. We have started adapting this dataset for the Russian language. The present paper analyzes the Visual Genome dataset, which is described as a comprehensive collection of visual data, detailing the creation of the dataset and the features of the resulting annotation. The process of translating the dataset into Russian is outlined and the challenges encountered and the methods used to address them are discussed. The results of the statistical analysis of the data obtained after translation are examined separately, the primary focus being on textual descriptions of images and objects. The paper identifies the primary reasons why different English phrases resulted in identical Russian translations, leading to shifts in data distribution compared to the original set. In summary, we present our reflections on the features of Visual Genome and the data obtained during the translation process. We then proceed to enumerate several potential applications for the translated data, accompanied by the path forward for further research.
The paper is devoted to the analysis and interpretation of multichannel electroencephalogram in the problem of recognizing the P300 component. We compare approaches based on statistical hypothesis testing and neural network models with trainable convolutions that can take into account the temporal structure of the signal. To identify informative differences between classes, the nonparametric Wilcoxon test and the Holm–Bonferroni correction at multiple comparisons are used. The most pronounced differences between target and nontarget responses were shown to be observed starting from approximately 200 ms after stimulus presentation, which is consistent with the physiology of evoked potentials. In this case, channel-oriented statistical analysis turned out to be less sensitive than deep learning models, which emphasizes the practical value of neural network methods for this problem. The obtained results can be used in the construction and interpretation of brain–computer interface systems and in the selection of informative time windows.
Regulatory requirements are currently presented as a rule in a semistructured format in human-readable documents, making the compliance checking of information models possible only manually. The first step toward automating compliance checking is the annotation of semantic components and domain-specific terms in the text of the requirements. However, such annotation requires significant time and deep semantic expertise from the user. The present study focuses on automating the annotation of regulatory requirements to simplify their translation into Web Ontology Language, thereby enabling subsequent automated compliance checking. We propose methods for automating annotation across three levels of the annotation scheme. These are domain-specific terms, semantic types, and semantic roles. The proposed approach demonstrates promising results: Recall@10 of 64.8
In this paper, an improved approach for automatic wildlife detection in natural environments based on the integration of a neural network architecture with a two-stream attention mechanism and a novel preclassification step based on infrared data has been presented. The proposed method addresses one of the key challenges in environmental monitoring: the need for scalable and accurate tools for assessing wildlife populations and supporting biodiversity conservation. The preclassification module analyzes color and thermal patterns in the infrared range to improve detection reliability before the main detection stage. This step made it possible to filter out background noise and select areas that are highly likely to contain living organisms, allowing for more focused and accurate subsequent analysis. The core detection system is built on a neural network architecture with a dual attention mechanism that emphasizes semantically significant areas of camera trap images while minimizing the influence of visually complex natural scenes. This is especially important for field data, which are often characterized by partial occlusions, uneven illumination, and rich backgrounds. Performance has been evaluated using the Wildlife Insights dataset, a large and diverse collection of camera trap images collected across different ecological regions. Based on experimental results, the proposed approach provides higher accuracy and robustness compared to traditional models, especially in visually complex environments. By combining infrared color-based preclassification and a detection pipeline enhanced by attention mechanisms, the proposed method significantly improves the efficiency of automated wildlife monitoring. The obtained results have confirmed the applicability of the system for ecological research, conservation planning, and long-term studies of wild animal populations.
Evaluation of the complexity of machine learning models is one of the main problems when selecting models for solving applied problems. The spectral properties of Hessian matrices make it possible to describe quantitatively the curvature of the optimization landscape and characterize the complexity of tuning the parameters of neural network models. In the study, a landscape measure of model complexity defined through the spectral properties of the Hessian matrices of the loss function and describing the change in landscape curvature when adding new data objects has been introduced. Based on theoretical estimates of the spectral norms of Hessian matrices, asymptotic estimates of the landscape complexity measure have been obtained for fully connected, 1D and 2D convolutional neural networks, as well as for convolutional networks with MaxPooling and AvgPooling operations, demonstrating an exponential dependence on the network depth and a polynomial dependence on architectural parameters. In the proposed approach, computationally feasible methods for estimating the landscape complexity of deep learning models based on analytical estimates of the spectral properties of Hessian matrices are provided.
This paper presents a method for multiclass segmentation of microscopic images of bacterial biofilms using deep learning. A comprehensive approach is proposed that includes overlapping patch-based processing, a procedure for fusing local predictions in overlapping regions, and semantic filtering of segmentation masks based on morphological features of the objects. The novelty of this work lies in combining segmentation, fusion, and filtering into a single system that ensures robustness to contrast inhomogeneity and textural characteristics of bacterial structures. The proposed approach eliminates patch-boundary artifacts and improves the consistency of the results. The developed solution is implemented as a web tool with interactive visualization and the ability to filter objects by morphological features. Experimental results confirm the effectiveness of the proposed method and its applicability for automated analysis of biofilm structures in biomedical research.
Cross-validation is one of the main methods for assessing the quality of solutions in machine learning. However, it remains unknown which of its varieties gives the most accurate results. To answer this question, it is necessary to learn how to calculate the variance of a cross-validation estimate. In this study, we will obtain some useful expressions for this quantity, which, in particular, make it possible to estimate the influence of the number of folds in splitting the sample. In the study, the effect of sample stratification during cross-validation has also been evaluated. To obtain a more pronounced difference, it has been proposed to perform the estimate using a procedure that is, in a sense, the opposite of stratification.
The paper discusses the application of a discrete apparatus to solve one of the central problems in machine learning, namely, the supervised classification problem. Special attention is paid to computationally complex procedures of searching for so-called maximum patterns in the training set. Traditional procedures of searching for these logical patterns have been improved, and the possibility of using neural networks for logical analysis and classification of big data has been demonstrated. The network architecture is based on an original methodology, according to which the problem of finding the maximum joint subsystem for a system of Boolean equations is solved at the training stage, the left-hand sides of the equations being represented by monotone conjunctive normal forms. The main feature of the implemented neural network is the presence of logical blocks “disjunction,” “conjunction,” and “negation.” The results of experiments on simulated and real data are given. The presented studies are a research priority and have important practical significance.
The special issue includes papers by the authors who presented the most interesting contributions closely related to the subject area of the journal. The papers can be divided into three groups: topics in recognition theory, studies in image processing, and studies in processing signals of other nature.
In the paper, the problem of searching for handwritten characters, glyphs, according to a specified pattern in a large file of handwritten text has been considered. This problem arises, in particular when searching for hieroglyphs in ancient Chinese handwritten texts when the search query is a digital image of the desired hieroglyph. The proposed search method is based on the generation of a feature description in the form of a small finite set of points called a constellation. The constellation is constructed based on the continuous skeleton of the query image. The search has been carried out based on an assessment of the similarity of constellations, the calculation of which is reduced to the selection of the best matching of points of two constellations by solving an optimization problem on assignments. The proposed method belongs to the class one-time search (one-shot search) since search is carried out using only one example at a time. Computational experiments have demonstrated high search quality comparable to advanced neural network methods. Furthermore, the constellation method does not require labeling and annotating large volumes of data, as well as a complex training procedure using powerful computing devices.
The application of generative models that use information retrieval, or the retrieval augmented generation approach, where a search engine retrieves relevant documents at the user’s request and the generative model produces an answer by combining the retrieved context with its internal knowledge, overcomes the limitations of large language models, such as knowledge obsolescence and the inability to handle dynamically changing data. The performance of such systems directly depends on the search quality, the context structure, and the wording of the instructions, but a comprehensive evaluation of this approach is a challenging task at present, not only due to the low number of suitable datasets and differences in the requirements for systems, but also because there are no standard metrics to measure the accuracy of the response in relation to the provided context. In this work, we conducted a comparative study of six language models (DeepSeek, Llama, Mistral, Qwen, RuAdapt, and YandexGPT) in a question-answering task using the retrieval augmented generation approach. The experiments were performed on four Russian-language datasets—XQuAD, TyDi QA, RuBQ, and SberQuAD—which were converted to a unified format suitable for the task. Various strategies for adding and arranging fragments in the context provided to the model were considered, and the evaluation of the models using the metrics of context relevance, utilization, completeness, adherence, and exact matching allowed us to identify the limitations of generative models in extracting information from relevant context.
Automated assessment of facial acne severity remains challenging due to subjective grading practices and limited consideration of lesion spatial characteristics. In this work, we address these challenges by first analyzing interexpert variability using a dataset independently annotated by three dermatologists and introducing a unified grading criterion to mitigate inconsistencies and systematic bias. Based on the harmonized labels, we conduct a comprehensive evaluation of deep learning approaches for continuous acne severity regression. We begin with strong baseline models that directly regress severity scores using convolutional and transformer-based architectures. We then explore detector-driven formulations that infer severity solely from lesion counts predicted by object detection models trained on a publicly available acne dataset, extending prior work by incorporating multiple modern detectors. To better align with clinical grading criteria, we further introduce heuristic geometric features that capture lesion coverage and spatial distribution across the face. Finally, we propose a hybrid framework that fuses global visual features extracted by a convolutional neural network with detector-derived lesion count and spatial features, jointly modeled using gradient boosting regression. Experimental results demonstrate that lesion count alone provides a competitive and interpretable signal, while incorporating spatial heuristics yields consistent improvements. The proposed fusion strategy achieves the best overall performance, outperforming standalone convolutional- or detector-based models. Additionally, we analyze the impact of label standardization and show that unified grading criteria reduce systematic bias in model predictions. These findings highlight the importance of combining global appearance cues with localized lesion characteristics for reliable and clinically meaningful acne severity assessment.
A mathematical apparatus of the queuing theory for simulating modern relational database management systems and, more broadly, distributed information systems is proposed. The proposed apparatus is shown to be applicable to simulating modern relational database management systems. A mathematical statement of Brewer’s theorem is proposed. Typical representatives of modern relational database management systems and test data that confirm strong consistency and availability of data with simultaneous data partition for them are considered. The problem of simulating a distributed database management system is stated. A set of models for calculating the basic parameters of a distributed system are proposed, as well as a set of queuing models suitable for simulating a distributed system. The parameters of the distributed system are linked to the terms of Brewer’s theorem so that one can confirm or refute its hypotheses during simulation. In further research, the mathematical models proposed in this article are planned to be refined, confirming their correctness and applicability to simulating distributed database management systems and information systems built using modern relational database management systems.
In this work, the effectiveness of machine learning algorithms with increased divergence in the prediction space for predicting the quantitative properties of inorganic compounds of ensembles is studied, including ensembles containing algorithms of different natures. The increased divergence is achieved by simultaneously adding algorithms to the ensemble with a maximum accuracy of approximation of the dependence and a maximum deviation from the algorithms previously included in the ensemble. The results of experiments have shown a reduction in the squared error for the ensembles with increased divergence, consisting of algorithms generated using the LightGBM method, by 18 to 66
Given datasets of various modalities and their ensemble for personal biometric identification, information-theoretic lower bounds on error probability and the corresponding upper bounds on accuracy as the functions of amount of processed information are investigated. The bounds are formed by the strictly monotonic functions that are the modifications of the well-known trade-off relations given by the rate-distortion function with the Hamming distortion measure. In the spaces of tree-structured and vector representations of face and signature images with given distances, the numerical realizations of the bounds are calculated for both the datasets of single modality objects and the ensemble of these datasets. The obtained realizations of the boundary relations demonstrate a possibility of increasing identification fidelity in the ensemble of datasets as well as more efficiency of the vector object representations. For datasets of the vector face and signature image representations and for their ensemble, the fidelity estimates of the decision algorithms on a basis of support vector machine (SVM) and random forest (RF) are calculated in terms of error probability and entropy of the decisions as amount of processed information. The obtained fidelity characteristics are compared with the appropriate boundary values.
