This study presents a comparative analysis of software efficiency for photogrammetric digitization and visualization of energy infrastructure objects. Experimental evaluation was conducted on a production boiler house section using Agisoft Metashape, 3DF Zephyr, Meshroom, RealityCapture, Pix4D, and the neural platform LumaAI. Results demonstrate that RealityCapture delivers superior reconstruction accuracy (1-10 mm error) and geometric detail preservation under complex reflective surface conditions, attributable to its hybrid data processing algorithms and GPU optimization. LumaAI exhibits rapid data processing capabilities and hidden area reconstruction technology (NeRF), but remains unsuitable for digitizing classified infrastructure due to data leakage risks. Critical limitations were identified for Meshroom (inefficiency with large frame sets) and Pix4D (inadaptability to terrestrial photogrammetry). Sanction-related deployment challenges for RealityCapture in the Russian Federation are highlighted. The findings substantiate the necessity for specialized domestic solutions integrating classical method precision with AI algorithms while ensuring cybersecurity. This research establishes fundamental software selection criteria for energy asset digitization, digital twin development, and VR simulator creation.

The paper considers the foreign experience of teaching engineering students to analyze speech signals using instrumental methods. Examples of obtaining spectrograms are given, as well as the capabilities of speech analysis software used in undergraduate and graduate engineering courses at the University of California (Los Angeles, USA). The disadvantages of speech analysis and visualization based on Fourier transform are shown. New solutions for processing and visualizing speech signals based on multilevel wavelet analysis are proposed. The main characteristics of the developed WaveView and WaveView-MWA programs that provide increased time-frequency resolution of vowel sounds are considered. For the first time, the results of high-precision analysis and visualization of consonant sounds - non-stationary signals inaccessible to spectral analysis using the Fourier transform are presented. A comparative analysis of the time-frequency resolution of spectrograms and wavelet sonograms in the visualization of English speech is performed. The developed technology of high-precision analysis and visualization of speech signals is used in the training of specialists of the Department of «Information Security» of the Faculty of «Informatics and Management» of the Bauman Moscow State Technical University during laboratory work on the course «Forensic study of phonograms».

The paper considers the principle of analytical transition to a local function at points on the domain of an implicit function defining a geometric object. Herewith, a transition to partial derivatives is provided to obtain a general form of an implicit local function describing the local geometry for any single point in the object domain. On the analogy of R-functional modeling, a mathematical apparatus for union/intersecting local geometric characteristics of a local function at a single point is provided to construct a discrete region of a complex geometric object. An example of the intersection of two functions on a defined domain of arguments demonstrates the obtaining of a discretely geometrized three-dimensional manifold for describing a cylinder. The proposed work is the continued development of method of the Functional Voxel Modeling which offers an analytical structure for the discrete-continuous description of complex geometric objects instead of the means of linear approximation currently used in this method.

This paper presents a method for the non-contact determination of the adiabatic wall temperature in high-speed gas flows. The method is based on the processing of a sequence of thermograms obtained using an IR camera, within a program developed in Python 3.10. The approach demonstrated high efficiency when handling large datasets, particularly concerning minimizing temporal and computational demands. The adiabatic wall temperature was determined under both steady-state conditions, directly in the experiment, and transient conditions, through the extrapolation of the heat flux as a function of the current temperature of the examined surface. The effectiveness of this method was demonstrated in the investigation of non-mechanical energy separation in compressible gas flows.

This article discusses ways to use visualization tools to build object classifiers during automation of a large enterprise. The proposed approaches allow stakeholders to get a visual representation and participate in the decisions required when building a classifier for large arrays of records. The use of visualization tools is considered when selecting classification objects, determining the attributes and values of classification attributes, ensuring the convenience of the classifier and implementing conflicting requirements from stakeholders. Among the proposed solutions, the methods of using system classes, building logical and physical models of the classifier, multidimensional classification, attribute-value data model, logical data model for describing the required analytics are described. The subject area is a classifier of works and services, examples of using the proposed solutions and the results of building a classifier at a large enterprise are given.

Single-pixel imaging is a method of computational imaging that allows to obtain images of objects using a photodetector that does not have spatial resolution. In this method, the object is illuminated by light having a special spatio-temporal structure, — light patterns, and a single-pixel photodetector measures the total amount of light reflected from the object. The possibility of obtaining an image and the image quality are closely related to the properties of the applied patterns and computational algorithms. In this paper, we consider patterns obtained from modified Hadamard matrices and study the features of image calculation using single-pixel imaging. We show the possibility of reducing both the sampling time and the computational resources required to obtain images by modifying the pattern system. The proposed theoretical method can be used in the practical implementation of the single-pixel imaging method in an experiment.

The article describes a new method for testing the independence of random data sets. This method uses representation of connections between data points in the form of nearest neighbor graphs and compares parameters of the resulting specific graph — such as the number of connected components and vertex degree distribution — to numerically derived critical statistics for random nearest neighbor graphs obtained by the author. The proposed method can be applied to various practical situations. It can be used to test the independence of random vectors in low-dimensional metric spaces. Such problems arise when analyzing data from physical measurements. Additionally, this approach is applicable for analyzing random points in high-dimensional spaces where direct numerical evaluations require exhaustive enumeration and are therefore impractical or impossible to obtain exactly. This problem relates to object classification characterized by many parameters. Moreover, proximity function between points may not necessarily be symmetric, which allows application of graph methods even in such cases. Alongside the task of sample testing, one could also consider comparing pseudo-random number generators by benchmarking structural graph statistics based on them. Considered probabilities of graph structure realization provide an independent set of criteria. For example, the number of fragments in some random graph might be typical for independent random variables while vertex degree distributions could differ significantly. This extends the applicability domain of statistical analysis. The paper presents a collection of model examples illustrating how the methodology works with respect to several types of practical scenarios mentioned above. A comparison of this method with other statistical approaches is provided. We emphasize that using graphs as visualization tools enables immediate identification of dependent elements within samples if there exist cluster centers represented by vertices with anomalously large degrees.

This paper examines the problem of targeted search for specific objects in a video stream on request and recording the timestamps of their appearance. Since the solutions currently available on the market were inadequate for the task, it was decided to implement such a tool independently as part of an ongoing research project conducted at the Keldysh Institute of Applied Mathematics of the Russian Academy of Sciences

This paper presents spectral rendering method that addresses key challenges in storing and processing spectral data. The proposed approach represents light and material properties using truncated Fourier coefficients, allowing spectra to be stored and manipulated compactly. This representation reduces memory usage and computational overhead while preserving the accuracy of spectral information during rendering. The method enables efficient reconstruction of stored spectral functions and simplifies operations such as color conversion. Several strategies for transforming Fourier coefficients within a path tracing framework are investigated, including different spectrum-to-color conversion techniques, such as using zeroth Fourier coefficient to directly convert Fourier-based spectrum to color. Experimental results show that the proposed method provides rendering quality comparable to traditional approaches while producing lower color noise and similar computation times. The method is particularly effective for fast preview and interactive rendering, where low samples per pixel are used and color noise strongly affects visual perception. Also, the paper describes applications of proposed method in neural rendering for storing BRDF using compact neural networks. Furthermore, variance reduction approach based on Fourier coefficients is proposed.

The generation of synthetic datasets for training neural models in object detection and recognition tasks has become a prevalent approach due to the cost and time savings compared to collecting real world data. However, synthetic images often lack critical details, which can degrade the performance of trained models. To address this issue, we propose the FakeSegment neural model, designed to annotate unrealistic parts of synthetic images. Our method utilizes two Single Shot Multibox Detector (SSD) networks with shared weights. By analyzing the differences in corresponding feature maps from real and images. By comparing these feature maps, we can pinpoint areas where synthetic images diverge from expected patterns observed in real-world data. FakeSegment automatically detects unnatural areas within the synthetic data. We evaluate our model on two datasets: the FantasticReality dataset and a newly introduced UnrealLanding dataset focused on aircraft safety during landing. Our contributions include (1) the development of the FakeSegment model, (2) the creation of the UnrealLanding dataset with paired synthetic and real images, and (3) a comprehensive evaluation demonstrating that Fake Segment outperforms baseline methods by 15% in Intersection over Union (IoU) for segmenting unreal parts.