Kirill Smelyakov

The object of research is the process of automated misinformation detection, with particular emphasis the use of large language models (LLMs) as scalable annotators to solve a problem of labeled data scarcity (typical for low-resource languages). Manual fact-checking and annotation are costly, slow, and difficult to scale, which limits the applicability of machine learning approaches. The research addresses the problem of enhancing the misinformation detection quality on small datasets by adding the LLM-generated synthetic labels to the training process and evaluates the achieved performance improvements. The results demonstrate that incorporating LLM-generated annotations into limited training data consistently improves classification performance. The observed gains range from approximately 2 to 3 percentage points in the F1-macro score depending on the configuration. The relative improvement is higher in scenarios with smaller manual data used. These findings can be explained by the increased volume and diversity of training data provided by synthetic annotations. LLMs provide good fact-checking across more data samples, enriching the dataset and helping the classifier generalize better. However, factors such as LLM version and size, inclusion of additional context, or the use of external knowledge retrieval influence annotation accuracy. In practical applications, the results suggest that LLM-based annotation can serve as an efficient and scalable method for bootstrapping small datasets. This approach is especially valuable for underrepresented languages, where annotated datasets and benchmarks are scarce, as it can increase misinformation detection accuracy without expensive manual annotation.

This article reviews current research on various approaches to improving images captured in low-light or dimly lit conditions. To address this issue, an adaptive ensemble method is proposed that combines several, specifically, four classical image processing algorithms and automatically distributes weights among them based on local quality. The developed system uses CLAHE, gamma correction, the Retinex algorithm, and automatic contrast stretching as its basic components. A comparative analysis of the developed method with the modern Deep Learning approach, Zero-DCE, was conducted on the real-world LOL dataset. The experimental results show that the proposed classical ensemble approach achieves a PSNR of 17.91 dB, outperforming the machine learning model by 14.2 percent, while the method does not require a training phase on datasets or GPU computations.

Colorizing monochromatic images is crucial for enhancing the informativeness of visual data in modern information technology and computer vision systems. However, automatic colorization poses an inherently ill-posed mathematical challenge due to the multimodal nature of color distributions, where multiple valid color mappings exist for any given grayscale input. This article conducts a comparative analysis of classical algorithms (such as the Welch and Levin methods) against advanced deep learning pipelines. The evaluated architectures range from baseline convolutional neural networks (CNNs) and standalone U-Net models to generative adversarial networks (GANs) and a novel hybrid Fusion GAN that integrates global semantic priors extracted via a ResNet-18 backbone. All models were trained and rigorously evaluated in the CIE Lab* color space, which effectively separates luminance from chrominance. To ensure a robust evaluation, experiments were conducted using a diverse benchmark of 2,000 images from the COCO dataset. Quality assessment combined traditional metrics like Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) with the Learned Perceptual Image Patch Similarity (LPIPS) metric, prioritizing human-like visual fidelity. Experimental results yielded counterintuitive insights. While the proposed Fusion GAN successfully surpassed classical methods, baseline CNNs, and standard GANs across most benchmarks, the standalone U-Net architecture secured the highest overall ranking. Specifically, U-Net achieved the top SSIM score of 0.945 (indicating superior structural preservation), the lowest LPIPS of 0.180 (best perceptual quality), and the fastest inference speed, enabling real-time applications.

The object of research is the process of designing and modeling intelligent agents of a neural network for the behavior of non-player characters (NPCs). Today, the field of video game development continues to progress, but the methodology for creating the “brain” of NPCs still uses primitive approaches with technological limitations. This creates the problem of ludonarrative dissonance, which actualizes this work, that offers a new solution. The research results demonstrate the potential of a new approach to creating intelligent and realistic NPCs in video games. The developed models demonstrate convergence of training, which confirms the ability to understand complex concepts, such as the psychological personality model International Personality Item Pool 50. The conclusions of the work are based on the results of theoretical analysis, technical modeling and practical experimentation. Analysis of the collected data forms the final confirmation of the feasibility of the proposed theory. An increase in the total step reward from –0.00275 to 0.03176 by the median for the best option and from –0.00733 to 0.02734 for the worst was recorded, taking into account the mathematical limit in [–1.0219; 0.9861] and aggressive normalization of the hyperbolic tangent function. The adaptation of models to the environmental conditions was confirmed by reducing the penalty by 30.47% by the median or by 23.59% by the extreme 5% of the data, for the worst configuration. The obtained results can be used in game development, to create the “brain” of NPCs. This will allow to get rid of the exponential complexity of development, inherent in current methodologies due to the proposed approach. Using reinforcement learning technologies, it is possible to make a more variable and detailed virtual context, with lower computational resources consumption, avoiding ludonarrative dissonance.

This paper presents a comparative analysis of the effectiveness of modern vector representation (embedding) methods for building a recommendation system within the online perfume retail segment. The primary objective was to investigate and compare various methods for constructing recommendation engines. The object of study is the decision-making process involved in providing recommendations in an online perfume store. The research focused on comparing three technological solutions: OpenAI models (text-embedding-3-small), Google Gemini (text-embedding-004), and the open-source Nomic Embed model, deployed locally via the Ollama platform. The experiment utilized a dataset containing over 8,000 perfume items, where each object was described through a structure of olfactory notes (top, middle, and base notes), brands, concentrations, and target audiences. The architecture of the recommendation system was developed and implemented using the Laravel framework integrated with vector search capabilities. A hybrid mechanism was realized, combining semantic similarity (Content-Based) with heuristic boosting based on brand and category. To calculate the similarity between products in the vector space, the cosine distance method was applied. System performance was evaluated using Precision, Recall, the F1-Score, and average response time. The results demonstrated that the OpenAI model provides the highest quality of semantic analysis, making it the superior choice for deep personalization. Meanwhile, the Ollama (Nomic) model emerged as the leader in processing speed, offering response times critical for high-load, real-time systems. The findings provide a balanced framework for selecting a model based on business priorities: maximum accuracy versus high performance with minimal infrastructure costs.

The subject matter of the study is the methods, algorithms, and software tools for annotating visual data in computer vision systems within the Data-Centric AI paradigm. The study analyzes the processes of structuring unstructured information where labeling quality determines the accuracy of artificial intelligence models, considering the limitations of manual methods and risks of systematic errors in full automation. The goal of the work is to perform a comprehensive analysis of modern image annotation methods and tools in computer vision systems and a quantitative evaluation of the effectiveness of hybrid Human-in-the-Loop strategies to improve the efficiency of forming high-quality datasets within the Data-Centric AI paradigm. The following tasks were addressed in the article: a systematization of annotation types from image classification to panoptic segmentation and 3D scenes was conducted; a review of the toolset based on the Segment Anything Model and GroundingDINO was performed; comparative evaluations of manual, automatic, and hybrid labeling scenarios based on accuracy (mIoU) and labor intensity were carried out; issues regarding operator trust and interaction ergonomics were identified. The following methods are used – systematic comparative analysis of hybrid Human-in-the-Loop strategies, cross-domain synthesis of active learning and interactive segmentation studies, and formalization of acceleration, quality, and manual labor indicators. The following results were obtained: the hybrid pipeline provides a 5.4x process acceleration. The formation time of a semantic mask for an object is reduced from 65.0 to 12.0 seconds while maintaining quality at mIoU = 0.94, with only a 0.02 loss relative to the reference standard. The hybrid scenario was found to be optimal within the threshold range of , covering the broadest class of practical tasks from training production models to medical diagnostics. A direct dependence of annotator performance on the reliability of automatic prompts was identified. Conclusions: hybrid annotation is the optimal strategy for creating Ground Truth in critical domains such as autonomous driving and medicine, providing a balance between speed and accuracy. The proposed formalization of the optimization problem with threshold constraint enables informed selection of the annotation scenario for a specific domain. Future research involves improving synthetic data generation to overcome the domain gap and developing adaptive interfaces to reduce cognitive load.

The subject of study is the linear dimensions of three-dimensional objects based on computed tomography data, specifically metallic foreign bodies in human tissues and organs. The purpose of the study is to develop a mathematical model and a method for automated, high-precision measurement of the linear dimensions of three-dimensional objects based on computed tomography results, with implementation in the form of software. Objectives: to formulate a mathematical model for representing a three-dimensional object in voxel space; to develop a method for segmenting metal fragments based on computed tomography results; to propose a method for spatial alignment of segmented objects based on principal component analysis; develop a method for determining the maximum linear dimension of an object in a new coordinate system; implement the proposed model and method as software and experimentally verify the measurement accuracy. Research methods: analysis of tomographic parameters, threshold segmentation with adaptive selection of threshold values, wave-based algorithm for finding connected components, principal component analysis to determine the object’s orientation, voxel modeling, and calculation of Euclidean distances between boundary points of a three-dimensional object. Results. A mathematical model for representing a three-dimensional object and a method for automated high-precision determination of its maximum linear dimension are proposed. The method was implemented as a software module and tested on 72 samples of metal fragments of six types embedded in the biological tissues of pig organs. The average deviation does not exceed 3%, and in the most complex cases remains within 5%, which indicates the high accuracy and stability of the proposed approach. Conclusions: The developed model and method ensure automated and objective determination of the linear dimensions of foreign bodies based on computed tomography data without operator intervention. The proposed software can be used in military medicine, forensic medical examination, disaster medicine, and healthcare facilities where the speed and reliability of diagnostic decisions are critically important.

This paper introduces a new way to describe fingerprint details using basic geometric properties to improve how well we can identify people biometrically. The method concentrates on where fingerprint lines end or split, fixing issues with older methods like SIFT and HOG, and avoiding the heavy load of deep learning. The descriptor uses distances and angles around the center of fingerprint details, making it resistant to changes in position, rotation, and size. The goals include checking how well the details stay the same, creating a model based on geometry, and testing it on regular data. By mixing math proofs with real-world adjustments, the model can spot shifts 92.85% of the time, which is helpful for aligning fingerprints, but needs extra work for negative matches. This simple, non-machine learning method can be easily expanded for different biometric uses.

A high-precision method for predicting the linear dimensions of three-dimensional objects of arbitrary shape based on computed tomography data has been developed. The method is aimed at reducing measurement errors caused by partial-volume effects, shadow transitions, and arbitrary spatial orientation of the object. The features of voxel image formation for metallic fragments within a tomographic volume and the influence of non-uniform density distribution on the accuracy of boundary determination have been investigated. A segmentation algorithm has been proposed that combines threshold-based selection of high-density voxels, their clustering into connected three-dimensional sets using a wave-based method, and construction of the object boundary surface. The procedure for evaluating geometric parameters has been improved by introducing a three-layer density-gradient distribution model and using additional points within each voxel to predict coordinates in the shadow region. It has been established that accounting for the partial contribution of the object within shadow voxels makes it possible to reconstruct its actual contours more accurately than with standard segmentation. It has been found that the use of principal component analysis eliminates the influence of fragment orientation in space and enables correct determination of its principal axis. It has been shown that the combination of segmentation, boundary prediction, and alignment along the principal axis improves the accuracy of estimating the linear dimensions of metallic fragments from CT data and ensures the suitability of the method for automated analysis. It has been established that the average relative error in determining the length of the studied samples does not exceed 5 %, and the obtained results confirm the practical feasibility of the proposed approach. Prospects for further research have been identified, consisting in adapting the method for automated decision support systems and extending its application to other types of tomographic data.

The aim of this work is to enhance the quality of visualization of computed tomography (CT) results by improving digital image processing methods, particularly through the use of the gradation correction (GC). The research methodology involves the application of algorithms for the automatic determination of the optimal pixel brightness range in CT images, which reduces the influence of extreme brightness values and increases the contrast of diagnostically significant structures. The proposed gradation correction is applied not within the standard 0–255 range but based on the statistical distribution of brightness, with the exclusion of non-informative tail regions of the histogram. The research results indicate that the proposed approach improves the visual clarity of organ boundaries and pathological formations without loss of diagnostic information and can be integrated into automated medical image analysis systems. The scientific novelty of the work lies in the enhancement of the gradation correction method for CT image processing through the automation of brightness range selection. The practical significance of the study is in improving the efficiency of visual analysis and computer-aided diagnostics at the stages of preliminary medical image processing, contributing to higher accuracy in pathology detection and reduced evaluation time.