M. G. Mozerov
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
M. G. Mozerov has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
- V. I. Kober
Kharkevich Institute for Information Transmission Problems, Russian Academy of Sciences, Moscow, Russia
18 shared publications - Viktor N. Karnaukhov
Kharkevich Institute for Information Transmission Problems, Russian Academy of Sciences, Moscow, Russia
18 shared publications - L. V. Zimina
Moscow Polytechnic University, Moscow, Russia
7 shared publications - Sergei M. Voronin
Chelyabinsk State University, Chelyabinsk, Russia
4 shared publications A. V. Voronin
Chelyabinsk State University, Chelyabinsk, Russia
3 shared publications- Artyom Makovetskii
Chelyabinsk State University, Chelyabinsk, Russia
3 shared publications
Latest Publications
- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 406-413
Scene Flow Estimation in the 3D Space Using a Locally Rigid Approach
Scene flow estimation algorithms generate a 3D field of motion of a dynamic scene. These methods are applied in robotics and autonomous driving. A common approach is to learn scene flow directly from 3D point clouds. Scene flow describes the temporal relationship between two successive point clouds. Scene flow is estimated by minimizing the distance between corresponding points in two point clouds. Despite achieving encouraging results, the point-matching methods used in these approaches typically ignore potentially structured point motion. A scene flow estimation algorithm using locally orthogonal transformations is proposed. To illustrate the efficiency of the proposed method, the results of the computer simulation are presented.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 406-413
Scene Flow Estimation in the 3D Space Using a Locally Rigid Approach
Scene flow estimation algorithms generate a 3D field of motion of a dynamic scene. These methods are applied in robotics and autonomous driving. A common approach is to learn scene flow directly from 3D point clouds. Scene flow describes the temporal relationship between two successive point clouds. Scene flow is estimated by minimizing the distance between corresponding points in two point clouds. Despite achieving encouraging results, the point-matching methods used in these approaches typically ignore potentially structured point motion. A scene flow estimation algorithm using locally orthogonal transformations is proposed. To illustrate the efficiency of the proposed method, the results of the computer simulation are presented.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 399-405
On the Reliability of Computer Vision Systems: Identifying Causal Mechanisms in Vision–Language Model Decision Making
Modern artificial intelligence architectures are becoming increasingly transparent, enabling more accurate assessment of model reliability and the development of robust training methodologies. Nevertheless, the problem of internal consistency in multimodal systems, particularly vision–language models (VLMs), remains unresolved. Notably, VLMs exhibit a pronounced shape bias, a preference for object shape over texture, mirroring aspects of human visual perception. An open question in the literature is whether this behavior reflects a fundamental property of visual representations or arises as an artifact of the language decoding process. In this work, we investigate the roles of shape and texture features within the internal layers of VLMs. Our methodology is based on linear probing, a diagnostic tool for assessing the linear separability of features in hidden representations, using the Stylized-ImageNet dataset. We identify a substantial gap between model behavior and internal representations: despite a strong shape-based preference at the output level (56–62%), texture information remains significantly more linearly separable across all layers (\({{F}_{1}} > 0.9\)) than geometric shape information (\({{F}_{1}} \approx 0.6\)). Moreover, targeted interventions in the hidden states of late layers result in highly unstable predictions. The results demonstrate that the observed shape bias is not driven by the dominance of geometric features in the internal representations but rather emerges during the language decoding stage. The proposed approach, grounded in mechanistic interpretability, elucidates the computational mechanisms underlying these models, thereby establishing a theoretical foundation for enhancing the predictability of computer vision systems in specialized applied scenarios.- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 399-405
On the Reliability of Computer Vision Systems: Identifying Causal Mechanisms in Vision–Language Model Decision Making
Modern artificial intelligence architectures are becoming increasingly transparent, enabling more accurate assessment of model reliability and the development of robust training methodologies. Nevertheless, the problem of internal consistency in multimodal systems, particularly vision–language models (VLMs), remains unresolved. Notably, VLMs exhibit a pronounced shape bias, a preference for object shape over texture, mirroring aspects of human visual perception. An open question in the literature is whether this behavior reflects a fundamental property of visual representations or arises as an artifact of the language decoding process. In this work, we investigate the roles of shape and texture features within the internal layers of VLMs. Our methodology is based on linear probing, a diagnostic tool for assessing the linear separability of features in hidden representations, using the Stylized-ImageNet dataset. We identify a substantial gap between model behavior and internal representations: despite a strong shape-based preference at the output level (56–62%), texture information remains significantly more linearly separable across all layers (\({{F}_{1}} > 0.9\)) than geometric shape information (\({{F}_{1}} \approx 0.6\)). Moreover, targeted interventions in the hidden states of late layers result in highly unstable predictions. The results demonstrate that the observed shape bias is not driven by the dominance of geometric features in the internal representations but rather emerges during the language decoding stage. The proposed approach, grounded in mechanistic interpretability, elucidates the computational mechanisms underlying these models, thereby establishing a theoretical foundation for enhancing the predictability of computer vision systems in specialized applied scenarios.- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 391-398
Neural Network Algorithm for Sampling 3D Point Clouds Using Deep Learning
Point cloud registration is the task of aligning two or more distinct point clouds by estimating the relative geometric transformation between them. This is a well-known problem, which plays an important role in many applications, including SLAM, 3D reconstruction, mapping, positioning, and localization. In this study, a method for point cloud sampling is proposed that is based on a neural network algorithm. The use of an autoencoder to the problem of aligning two point clouds makes it possible to determine a set of points that can be removed from both clouds without significant reduction of the registration accuracy. This problem is relevant because the cloud dimensionality significantly affects the registration algorithm rate. To illustrate the performance of the proposed method, the results of the computer simulation are presented.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 71 (6) , 391-398
Neural Network Algorithm for Sampling 3D Point Clouds Using Deep Learning
Point cloud registration is the task of aligning two or more distinct point clouds by estimating the relative geometric transformation between them. This is a well-known problem, which plays an important role in many applications, including SLAM, 3D reconstruction, mapping, positioning, and localization. In this study, a method for point cloud sampling is proposed that is based on a neural network algorithm. The use of an autoencoder to the problem of aligning two point clouds makes it possible to determine a set of points that can be removed from both clouds without significant reduction of the registration accuracy. This problem is relevant because the cloud dimensionality significantly affects the registration algorithm rate. To illustrate the performance of the proposed method, the results of the computer simulation are presented.
- 1 March 2025
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 70 (3) , 104-110
Estimation of 3D Point Cloud Coordinates Using Stereo Matching and Superpixel Segmentation
An important element of the movement of robotic systems is the determination of the object’s position in the coordinates associated with the scene, as well as the vector of the autonomous device’s own movement in this coordinate system (rotation matrix and displacement vector). All these parameters can be determined by estimating the values of the 3D coordinates of the cloud of scene points associated with the coordinate system of the onboard video camera of an autonomous robotic device. The process of finding the rotation matrix and the displacement vector is called the registration of three-dimensional point clouds and is of great importance in robotics and computer vision. Therefore, there are many methods of such registration. However, the problem is how to estimate the values of the 3D coordinates of the point cloud of the scene. In this paper, we propose to determine the desired coordinates of points using stereomatching. The main assumption of classical stereomatching is that matching pixels in stereo images have the same brightness values. However, this assumption is generally incorrect, given the presence of noise in the images and the different illumination of the left and right images in the stereo pair. In addition, pixel-by-pixel mapping does not work in areas where there is no real texture. The methods that solve all the problems of dense stereo patching are quite complex and require high computational costs. However, the problem can be solved in a simpler and less costly way, given that the clouds intended for registration consist of a small number of points, significantly smaller than the pixels in the image. Therefore, stereomatching in our case is carried out based on a certain neighborhood of the matched pixels, called a superpixel. The accuracy of such matching surpasses the use of a feature close to the principle of robust sensation. In addition, we have proposed an original method for verifying the reliability of the disparity determination based on local autocorrelation, thus removing cloud points with unreliable disparity estimates. As a result, our estimate of the 3D coordinates of “reliable” scene points is comparable to the best state-of-the-art deep learning algorithms.
- 1 March 2025
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 70 (3) , 104-110
Estimation of 3D Point Cloud Coordinates Using Stereo Matching and Superpixel Segmentation
An important element of the movement of robotic systems is the determination of the object’s position in the coordinates associated with the scene, as well as the vector of the autonomous device’s own movement in this coordinate system (rotation matrix and displacement vector). All these parameters can be determined by estimating the values of the 3D coordinates of the cloud of scene points associated with the coordinate system of the onboard video camera of an autonomous robotic device. The process of finding the rotation matrix and the displacement vector is called the registration of three-dimensional point clouds and is of great importance in robotics and computer vision. Therefore, there are many methods of such registration. However, the problem is how to estimate the values of the 3D coordinates of the point cloud of the scene. In this paper, we propose to determine the desired coordinates of points using stereomatching. The main assumption of classical stereomatching is that matching pixels in stereo images have the same brightness values. However, this assumption is generally incorrect, given the presence of noise in the images and the different illumination of the left and right images in the stereo pair. In addition, pixel-by-pixel mapping does not work in areas where there is no real texture. The methods that solve all the problems of dense stereo patching are quite complex and require high computational costs. However, the problem can be solved in a simpler and less costly way, given that the clouds intended for registration consist of a small number of points, significantly smaller than the pixels in the image. Therefore, stereomatching in our case is carried out based on a certain neighborhood of the matched pixels, called a superpixel. The accuracy of such matching surpasses the use of a feature close to the principle of robust sensation. In addition, we have proposed an original method for verifying the reliability of the disparity determination based on local autocorrelation, thus removing cloud points with unreliable disparity estimates. As a result, our estimate of the 3D coordinates of “reliable” scene points is comparable to the best state-of-the-art deep learning algorithms.
- 1 January 2024
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 69 (1-3) , 29-37
Occlusion Handling in Depth Estimation of a Scene from a Given Light Field Using a Geodesic Distance and the Principle of Symmetry of the View
The problem of depth estimation of a scene from a given light field can be reduced to the problem of classical stereo matching with the statement that matching pixels in stereo images have the same brightness values. However, this assumption is generally incorrect, given the presence of noise in the images and the different illumination of the left and right images in the stereo pair, as well as the existence of occlusions. In this regard, the luminous flux, representing 80 images pairwise epipolar to the central one, offers a number of advantages, especially in terms of occlusion handling. In this paper, we propose the principle of viewing symmetry: if a pixel of the central image lies in the occlusion zone relative to one of the peripheral images of the light field, then this pixel does not belong to the occlusion zone for an axisymmetric image of the peripheral field. Thus, it is possible to form a robust volume of discrepancy weights relative to the occlusion. As a result, the algorithm proposed in the article significantly improves the result of the scene depth reconstruction. The effectiveness of our approach is demonstrated using the main test database of the light field and comparing it with the best reconstruction algorithms in the efficiency of border recognition and in the speed of calculation.
- 1 January 2024
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 69 (1-3) , 29-37
Occlusion Handling in Depth Estimation of a Scene from a Given Light Field Using a Geodesic Distance and the Principle of Symmetry of the View
The problem of depth estimation of a scene from a given light field can be reduced to the problem of classical stereo matching with the statement that matching pixels in stereo images have the same brightness values. However, this assumption is generally incorrect, given the presence of noise in the images and the different illumination of the left and right images in the stereo pair, as well as the existence of occlusions. In this regard, the luminous flux, representing 80 images pairwise epipolar to the central one, offers a number of advantages, especially in terms of occlusion handling. In this paper, we propose the principle of viewing symmetry: if a pixel of the central image lies in the occlusion zone relative to one of the peripheral images of the light field, then this pixel does not belong to the occlusion zone for an axisymmetric image of the peripheral field. Thus, it is possible to form a robust volume of discrepancy weights relative to the occlusion. As a result, the algorithm proposed in the article significantly improves the result of the scene depth reconstruction. The effectiveness of our approach is demonstrated using the main test database of the light field and comparing it with the best reconstruction algorithms in the efficiency of border recognition and in the speed of calculation.
- 1 January 2024
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 69 (1-3) , 21-28
Superpixel-Segmentation Based on Energy Minimization and Convolution with the Geodesic Distance Kernel
—The energy minimization or maximum a posteriori probability (MAP) method is the basis for solving many computer vision problems, including the segmentation problem. However, it is assumed that the number of regions during segmentation is quite small. At the same time, in the problem of superpixel segmentation or otherwise excessive segmentation, the number of such areas exceeds 1000, which makes the computational optimization problem by the MAP method practically impossible. In this paper, we propose a solution that reduces segmentation with any number of areas to the problem of marking only nine labels. In addition, convolution with the geodesic distance kernel is used to enhance the robustness of optimization. This makes it possible to obtain single-linked superpixels at the output of the algorithm, unlike many other methods that require additional adjustments. The effectiveness of the proposed method is compared and measured by the precision-recall criteria, as well as by visual illustration.
- 1 January 2024
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 69 (1-3) , 21-28
Superpixel-Segmentation Based on Energy Minimization and Convolution with the Geodesic Distance Kernel
—The energy minimization or maximum a posteriori probability (MAP) method is the basis for solving many computer vision problems, including the segmentation problem. However, it is assumed that the number of regions during segmentation is quite small. At the same time, in the problem of superpixel segmentation or otherwise excessive segmentation, the number of such areas exceeds 1000, which makes the computational optimization problem by the MAP method practically impossible. In this paper, we propose a solution that reduces segmentation with any number of areas to the problem of marking only nine labels. In addition, convolution with the geodesic distance kernel is used to enhance the robustness of optimization. This makes it possible to obtain single-linked superpixels at the output of the algorithm, unlike many other methods that require additional adjustments. The effectiveness of the proposed method is compared and measured by the precision-recall criteria, as well as by visual illustration.
- 1 December 2023
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 33 (4) , 1242-1249
Development of Computer Vision, Image Processing, and Analysis at the Digital Optics Laboratory of the Institute for Information Transmission Problems of the Russian Academy of Sciences
Optical and digital images are one of the most important channels for transmitting information. At the Institute for Information Transmission Problems of the Russian Academy of Sciences (IITP RAS), this topic has always, since the founding of the institute, been given the closest attention. The institute’s employees have made significant contributions to both the domestic and global science of image processing. In this work, the authors touched only on the main stages of the development of the theory, methods, and algorithms for image processing and analysis in the Laboratory of Digital Optics of the Institute for Information Transmission Problems of the Russian Academy of Sciences.
- 1 December 2023
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 33 (4) , 1242-1249
Development of Computer Vision, Image Processing, and Analysis at the Digital Optics Laboratory of the Institute for Information Transmission Problems of the Russian Academy of Sciences
Optical and digital images are one of the most important channels for transmitting information. At the Institute for Information Transmission Problems of the Russian Academy of Sciences (IITP RAS), this topic has always, since the founding of the institute, been given the closest attention. The institute’s employees have made significant contributions to both the domestic and global science of image processing. In this work, the authors touched only on the main stages of the development of the theory, methods, and algorithms for image processing and analysis in the Laboratory of Digital Optics of the Institute for Information Transmission Problems of the Russian Academy of Sciences.
- 1 December 2023
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 68 (S2) , S254-S262
Accurate and Fast Geodesic Distance Calculation Algorithm for Superpixel Segmentation
—Modeling in an affine space on the basis of a geodesic distance makes it possible to implement important computer vision techniques. Among these applications is superpixel segmentation, in which geodesic distances from the center of specified segments to an arbitrary image point are calculated. Meanwhile, the algorithms proposed so far for calculating such distances in segmentation problems have been heuristic, iterative approaches, which do not guarantee the expected result. In this study, a new fast algorithm for calculating the geodesic distance is proposed, which is proven to be accurate. The image segments obtained using this algorithm are simply connected regions. The algorithm yields simply connected superpixels at the output, in contrast to many other approaches based on the spatial proximity of the geodesic distance and requiring an additional correction. The proposed technique surpasses its analogs in the border recognition efficiency and computational speed.
- 1 December 2023
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 68 (S2) , S254-S262
Accurate and Fast Geodesic Distance Calculation Algorithm for Superpixel Segmentation
—Modeling in an affine space on the basis of a geodesic distance makes it possible to implement important computer vision techniques. Among these applications is superpixel segmentation, in which geodesic distances from the center of specified segments to an arbitrary image point are calculated. Meanwhile, the algorithms proposed so far for calculating such distances in segmentation problems have been heuristic, iterative approaches, which do not guarantee the expected result. In this study, a new fast algorithm for calculating the geodesic distance is proposed, which is proven to be accurate. The image segments obtained using this algorithm are simply connected regions. The algorithm yields simply connected superpixels at the output, in contrast to many other approaches based on the spatial proximity of the geodesic distance and requiring an additional correction. The proposed technique surpasses its analogs in the border recognition efficiency and computational speed.
- 19 December 2022
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 67 (12) , 1557-1563
Impainting by Restoring Image Gradients
Impainting is used both to improve the visual perception of an image and in classical problems of recognition and robotics to remove irrelevant information from an image. Modern impainting methods of reconstruction drawing use neural networks. However, these approaches are plagued with disadvantages that prevent these algorithms from being used in practical applications of computer vision. The authors have proposed in their recent paper an impainting method based on a non-local mean (NLM) filter that outperforms other conventional image processing methods in terms of peak signal-to-noise ration (PSNR). However, this algorithm may create artifacts at the boundaries of the rendering zones that distort visual perception. In this article, we propose to draw not the image, but the gradients of the image restored and only then restore the image itself, maintaining in this way connectivity at the border of the rendering areas. This method eliminates most of the artifacts that arise when using the previously proposed method, and also surpasses it in terms of the recovery accuracy assessed using the PSNR criterion.
- 19 December 2022
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 67 (12) , 1557-1563
Impainting by Restoring Image Gradients
Impainting is used both to improve the visual perception of an image and in classical problems of recognition and robotics to remove irrelevant information from an image. Modern impainting methods of reconstruction drawing use neural networks. However, these approaches are plagued with disadvantages that prevent these algorithms from being used in practical applications of computer vision. The authors have proposed in their recent paper an impainting method based on a non-local mean (NLM) filter that outperforms other conventional image processing methods in terms of peak signal-to-noise ration (PSNR). However, this algorithm may create artifacts at the boundaries of the rendering zones that distort visual perception. In this article, we propose to draw not the image, but the gradients of the image restored and only then restore the image itself, maintaining in this way connectivity at the border of the rendering areas. This method eliminates most of the artifacts that arise when using the previously proposed method, and also surpasses it in terms of the recovery accuracy assessed using the PSNR criterion.
- 19 December 2022
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 67 (12) , 1531-1541
Segmentation of Breast Masses in Digital Mammography Based on U-Net Deep Convolutional Neural Networks
In this paper, we propose a new algorithm for segmentation of breast masses in digital mammography based on U-Net deep convolutional neural networks. All raw mammogram images are preprocessed to improve segmentation reliability. Deep learning was carried out using additional methods of data augmentation and fine tuning of the neural network. The performance of the proposed algorithm for segmentation of breast masses using the known CBIS-DDSM database is discussed.
- 19 December 2022
- Published by Pleiades Publishing Ltd in Journal of Communications Technology and Electronics
- Vol. 67 (12) , 1531-1541
Segmentation of Breast Masses in Digital Mammography Based on U-Net Deep Convolutional Neural Networks
In this paper, we propose a new algorithm for segmentation of breast masses in digital mammography based on U-Net deep convolutional neural networks. All raw mammogram images are preprocessed to improve segmentation reliability. Deep learning was carried out using additional methods of data augmentation and fine tuning of the neural network. The performance of the proposed algorithm for segmentation of breast masses using the known CBIS-DDSM database is discussed.