Valentin Malykh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Valentin Malykh has a total of 78 co-authors. Here are the top co-authors by number of shared publications:
- Aleksei Samarin
ITMO University, St. Petersburg, Russian Federation
24 shared publications - Alexander Motyko
ITMO University, St. Petersburg, Russian Federation
23 shared publications - Elena Mikhailova
ITMO University, St. Petersburg, Russian Federation
22 shared publications - Alexander Savelev
ITMO University, St. Petersburg, Russian Federation
21 shared publications - Aleksei Toropov
ITMO University, St. Petersburg, Russian Federation
19 shared publications - Artem Nazarenko
ITMO University, St. Petersburg, Russian Federation
19 shared publications
Latest Publications
- Open Access
- 6 November 2026
- Published by MDPI AG in Technologies
- Vol. 14 (11) , 731
MFL-ScrollNet: Recurrent Sliding-Window Detection for Panoramic Pipeline Magnetic Flux Leakage Images
Automated interpretation of magnetic flux leakage (MFL) data is central to condition assessment of long-distance oil and gas pipelines. Yet, full-resolution panoramas are difficult to process without either suppressing small anomalies through resizing or fragmenting elongated objects at sliding-window boundaries. We propose MFL-ScrollNet, a recurrent sliding-window detector that formulates panoramic analysis as an ordered sequence problem. A gated recurrent unit propagates a compact hidden state across narrow axial windows, while geometric cross-window association in global panorama coordinates, complemented by learned novelty prediction, consolidates detections across adjacent windows. Experiments on the public PipeMFL-240K benchmark show that the full model attains an mAP50 of 0.535, an mAP50:95 of 0.332, a macro-averaged F1 score of 0.662, and a micro-averaged F1 score of 0.547. MFL-ScrollNet achieves competitive average-precision performance, surpassing the strongest reproduced static baseline, DINO-R50, while approaching the performance of the domain-specific STPDNet. At the selected operating point, MFL-ScrollNet achieves the highest macro- and micro-averaged F1 scores among the evaluated methods, demonstrating improved detection performance across both individual classes and the overall defect population. Controlled ablations show consistent gains from recurrent context, global-coordinate association, and learned novelty prediction. These results demonstrate that sequential cross-window reasoning improves both class-balanced and instance-level detection in panoramic MFL data.
- 6 November 2026
- Published by MDPI AG in Technologies
- Vol. 14 (11) , 731
Open AccessMFL-ScrollNet: Recurrent Sliding-Window Detection for Panoramic Pipeline Magnetic Flux Leakage Images
Automated interpretation of magnetic flux leakage (MFL) data is central to condition assessment of long-distance oil and gas pipelines. Yet, full-resolution panoramas are difficult to process without either suppressing small anomalies through resizing or fragmenting elongated objects at sliding-window boundaries. We propose MFL-ScrollNet, a recurrent sliding-window detector that formulates panoramic analysis as an ordered sequence problem. A gated recurrent unit propagates a compact hidden state across narrow axial windows, while geometric cross-window association in global panorama coordinates, complemented by learned novelty prediction, consolidates detections across adjacent windows. Experiments on the public PipeMFL-240K benchmark show that the full model attains an mAP50 of 0.535, an mAP50:95 of 0.332, a macro-averaged F1 score of 0.662, and a micro-averaged F1 score of 0.547. MFL-ScrollNet achieves competitive average-precision performance, surpassing the strongest reproduced static baseline, DINO-R50, while approaching the performance of the domain-specific STPDNet. At the selected operating point, MFL-ScrollNet achieves the highest macro- and micro-averaged F1 scores among the evaluated methods, demonstrating improved detection performance across both individual classes and the overall defect population. Controlled ablations show consistent gains from recurrent context, global-coordinate association, and learned novelty prediction. These results demonstrate that sequential cross-window reasoning improves both class-balanced and instance-level detection in panoramic MFL data.
- Open Access
- 26 October 2026
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 10 (11) , 363
PACTraD: Multi-Task Benchmark and Domain-Specific Attention Mechanism for Wildlife Classification, Detection, and Segmentation in Protected Area Camera Trap Data
Camera traps are a key tool for non-invasive wildlife monitoring, yet models trained on existing benchmarks often generalize poorly to new regions due to domain shift, class imbalance, and natural variability in illumination, season, and background. We introduce PACTraD, a benchmark of 116,973 annotated camera trap images collected in a protected area of the Baikal region across eight classes (deer, no_animals, bear, boar, hare, badger, lynx, wolf), with a pronounced long-tailed distribution (deer: 73.1%; lynx and wolf: 0.3% and 0.2%) and substantial seasonal and lighting variability, including infrared night imagery. The benchmark supports three tasks under a unified, reproducible protocol: classification, object detection, and semantic segmentation. We further propose a domain-specific attention mechanism augmenting self-attention with eleven learnable priors capturing contour, foreground/background separation, bilateral symmetry, discriminative part regions, texture frequency, confidence weighting, motion consistency, gradient enhancement, scale, color invariance, and shape compactness. Evaluated on six backbones for classification and detection and five for segmentation, against no-attention baselines and three generic attention modules (SE, SA, CBAM), it consistently performs best, raising weighted F1 for classification from 0.943 to 0.964 (DINOv3), AP50 for detection from 0.876 to 0.903 (InternImage-H), and mIoU for segmentation from 0.805 to 0.834 (SegMAN), with component-level ablations revealing task-specific contributions of each prior. However, performance on rare species remains limited (macro F1: 0.829 for classification), indicating that further research is needed to address long-tail recognition in camera trap data.
- 26 October 2026
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 10 (11) , 363
Open AccessPACTraD: Multi-Task Benchmark and Domain-Specific Attention Mechanism for Wildlife Classification, Detection, and Segmentation in Protected Area Camera Trap Data
Camera traps are a key tool for non-invasive wildlife monitoring, yet models trained on existing benchmarks often generalize poorly to new regions due to domain shift, class imbalance, and natural variability in illumination, season, and background. We introduce PACTraD, a benchmark of 116,973 annotated camera trap images collected in a protected area of the Baikal region across eight classes (deer, no_animals, bear, boar, hare, badger, lynx, wolf), with a pronounced long-tailed distribution (deer: 73.1%; lynx and wolf: 0.3% and 0.2%) and substantial seasonal and lighting variability, including infrared night imagery. The benchmark supports three tasks under a unified, reproducible protocol: classification, object detection, and semantic segmentation. We further propose a domain-specific attention mechanism augmenting self-attention with eleven learnable priors capturing contour, foreground/background separation, bilateral symmetry, discriminative part regions, texture frequency, confidence weighting, motion consistency, gradient enhancement, scale, color invariance, and shape compactness. Evaluated on six backbones for classification and detection and five for segmentation, against no-attention baselines and three generic attention modules (SE, SA, CBAM), it consistently performs best, raising weighted F1 for classification from 0.943 to 0.964 (DINOv3), AP50 for detection from 0.876 to 0.903 (InternImage-H), and mIoU for segmentation from 0.805 to 0.834 (SegMAN), with component-level ablations revealing task-specific contributions of each prior. However, performance on rare species remains limited (macro F1: 0.829 for classification), indicating that further research is needed to address long-tail recognition in camera trap data.
- 14 October 2026
- Published by Springer Nature
- p. 147-162
Beyond Attention: Module-Level Refusal Directions in Mamba-Based Hybrid Language Models
- 14 October 2026
- Published by Springer Nature
- p. 147-162
Beyond Attention: Module-Level Refusal Directions in Mamba-Based Hybrid Language Models
- 14 October 2026
- Published by Springer Nature
- p. 212-226
A Retrieval-Augmented Pipeline for Hittite-to-English Translation from Transliteration
- 14 October 2026
- Published by Springer Nature
- p. 212-226
A Retrieval-Augmented Pipeline for Hittite-to-English Translation from Transliteration
- Open Access
- 14 October 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (10) , 330
A Reproducible Microscopy Acquisition Pipeline and Microscopy-Aware Attention for Classification, Detection, and Segmentation of Microorganisms
Automated microscopic image analysis is essential for quantitative microbiology, yet manual acquisition and generic recognition models struggle with morphological variability and scene artifacts. We address both ends of this pipeline. First, we present a low-cost, open-source acquisition system, based on an Arduino UNO R3 microcontroller and NEMA 17 stepper motors, that automates focusing, magnification, and illumination on microscopes from the Levenhuk MED series, including the Levenhuk MED 20T, MED 25T, MED 30T, MED 35T, and MED 40T, and applies a standardized, programmable control procedure intended to improve the reproducibility of imaging conditions. Using this system, we construct a benchmark of 100,000 microscopy images across five classes (micrococci, diplococci, streptococci, bacilli, and background), with annotations for classification, detection, and segmentation. Second, we propose Microscopy-Aware Attention (MAA), which augments standard self-attention with six learnable, biologically motivated priors capturing cell boundaries, focus quality, internal-structure contrast, spatial compactness, shape, and local density. MAA is evaluated on six classification backbones, six detection backbones, and five segmentation backbones against no-attention, Squeeze-and-Excitation, spatial-attention, and CBAM baselines. It consistently yields the best results, raising weighted F1 for classification from 0.947 to 0.975 (DINOv3), for detection from 0.901 to 0.928 (InternImage-H), and mIoU for segmentation from 0.834 to 0.862 (SegMAN), with component-wise ablations showing task-specific contributions of each prior.- 14 October 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (10) , 330
Open AccessA Reproducible Microscopy Acquisition Pipeline and Microscopy-Aware Attention for Classification, Detection, and Segmentation of Microorganisms
Automated microscopic image analysis is essential for quantitative microbiology, yet manual acquisition and generic recognition models struggle with morphological variability and scene artifacts. We address both ends of this pipeline. First, we present a low-cost, open-source acquisition system, based on an Arduino UNO R3 microcontroller and NEMA 17 stepper motors, that automates focusing, magnification, and illumination on microscopes from the Levenhuk MED series, including the Levenhuk MED 20T, MED 25T, MED 30T, MED 35T, and MED 40T, and applies a standardized, programmable control procedure intended to improve the reproducibility of imaging conditions. Using this system, we construct a benchmark of 100,000 microscopy images across five classes (micrococci, diplococci, streptococci, bacilli, and background), with annotations for classification, detection, and segmentation. Second, we propose Microscopy-Aware Attention (MAA), which augments standard self-attention with six learnable, biologically motivated priors capturing cell boundaries, focus quality, internal-structure contrast, spatial compactness, shape, and local density. MAA is evaluated on six classification backbones, six detection backbones, and five segmentation backbones against no-attention, Squeeze-and-Excitation, spatial-attention, and CBAM baselines. It consistently yields the best results, raising weighted F1 for classification from 0.947 to 0.975 (DINOv3), for detection from 0.901 to 0.928 (InternImage-H), and mIoU for segmentation from 0.834 to 0.862 (SegMAN), with component-wise ablations showing task-specific contributions of each prior. - Open Access
- 11 August 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (8) , 236
Flow-Guided Neural Pruning: Signal-Flow Framework for Multi-Architecture Model Compression
This paper presents a novel method for pruning deep neural networks based on the concept of flow, derived from the continuous modeling of signal propagation across layers. We derive flow functions for fully connected, convolutional, and self-attention architectures, and we propose a new iterative pruning algorithm, Iterative Flow-Aware Pruning (IFAP), that leverages these measures to identify and eliminate non-essential parameters while preserving critical information pathways. Extensive experiments across ten prominent architectures (including CNNs, vision transformers, and efficient mobile networks) on ten benchmark datasets demonstrate consistent accuracy–compression trade-offs: of the evaluated configurations achieve a 60– reduction in computational cost relative to the corresponding baseline model. Furthermore, of the evaluated configurations retain more than of their baseline Top-1 accuracy. These results validate flow-based importance scoring as a robust and general-purpose foundation for model optimization.- 11 August 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (8) , 236
Open AccessFlow-Guided Neural Pruning: Signal-Flow Framework for Multi-Architecture Model Compression
This paper presents a novel method for pruning deep neural networks based on the concept of flow, derived from the continuous modeling of signal propagation across layers. We derive flow functions for fully connected, convolutional, and self-attention architectures, and we propose a new iterative pruning algorithm, Iterative Flow-Aware Pruning (IFAP), that leverages these measures to identify and eliminate non-essential parameters while preserving critical information pathways. Extensive experiments across ten prominent architectures (including CNNs, vision transformers, and efficient mobile networks) on ten benchmark datasets demonstrate consistent accuracy–compression trade-offs: of the evaluated configurations achieve a 60– reduction in computational cost relative to the corresponding baseline model. Furthermore, of the evaluated configurations retain more than of their baseline Top-1 accuracy. These results validate flow-based importance scoring as a robust and general-purpose foundation for model optimization. - Open Access
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 5198-5203
SLeDoC: System for Legal Document Comparison
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 5198-5203
Open AccessSLeDoC: System for Legal Document Comparison
- 1 June 2026
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 36 (2) , 323-334
Improving the Accuracy of Automatic Wildlife Detection in Nature Reserves Using Infrared Imaging
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.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 36 (2) , 323-334
Improving the Accuracy of Automatic Wildlife Detection in Nature Reserves Using Infrared Imaging
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.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 36 (2) , 302-312
Improvement in Microbial Classification Quality Using Synthetic Microscopic Images Generated by Large Visual-Language Models
The lack of annotated microscopic datasets remains a major obstacle to training robust deep learning models for microbial classification. In this paper, a novel data augmentation pipeline that uses visual–linguistic large-scale models to generate synthetic microscopic images of six different bacterial and nonbacterial classes has been proposed. Synthetic samples have gradually been added to the training set in controlled proportions in order to evaluate systematically their impact on the quality of the model. Quantitative experiments have shown that including up to 11% synthetic data leads to a significant increase in classification accuracy: the best configuration achieves a Precision of 0.91, a Recall of 0.89, and an F1-score of 0.90. However, when this saturation threshold is exceeded, a decrease in quality is observed, indicating possible introduction of distribution noise or overfitting due to excessive synthetic augmentation. The results have demonstrated the potential of generating synthetic data based on visual–linguistic large-scale models as a scalable approach to addressing class imbalance and data shortage issues in microbiological image analysis problems.
- 1 June 2026
- Published by Pleiades Publishing Ltd in Pattern Recognition and Image Analysis
- Vol. 36 (2) , 302-312
Improvement in Microbial Classification Quality Using Synthetic Microscopic Images Generated by Large Visual-Language Models
The lack of annotated microscopic datasets remains a major obstacle to training robust deep learning models for microbial classification. In this paper, a novel data augmentation pipeline that uses visual–linguistic large-scale models to generate synthetic microscopic images of six different bacterial and nonbacterial classes has been proposed. Synthetic samples have gradually been added to the training set in controlled proportions in order to evaluate systematically their impact on the quality of the model. Quantitative experiments have shown that including up to 11% synthetic data leads to a significant increase in classification accuracy: the best configuration achieves a Precision of 0.91, a Recall of 0.89, and an F1-score of 0.90. However, when this saturation threshold is exceeded, a decrease in quality is observed, indicating possible introduction of distribution noise or overfitting due to excessive synthetic augmentation. The results have demonstrated the potential of generating synthetic data based on visual–linguistic large-scale models as a scalable approach to addressing class imbalance and data shortage issues in microbiological image analysis problems.
- Open Access
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-28
IDAP++: Advancing Divergence-Aware Pruning with Joint Filter and Layer Optimization
Modern knowledge and large volumes of data are increasingly encoded within neural networks, making the task of simplifying their structures and reducing the number of parameters especially relevant, both to improve efficiency and to facilitate deployment in resource-constrained environments. This paper presents a novel approach to neural network compression that addresses redundancy at both the filter and architectural levels through a unified framework grounded in information flow analysis. Building upon the concept of tensor flow divergence, which quantifies how information transforms across network layers, we develop a two-stage optimization process. The first stage employs iterative divergence-aware pruning to identify and remove redundant filters while preserving critical information pathways. The second stage extends this principle to higher-level architecture optimization by analyzing layer-wise contributions to information propagation and selectively eliminating entire layers that demonstrate minimal impact on network performance. The proposed method naturally adapts to diverse architectures, including convolutional networks, transformers, and hybrid designs, providing a consistent metric for comparing the structural importance across different layer types. Experimental validation across multiple modern architectures and datasets reveals that this combined approach achieves substantial model compression while maintaining competitive accuracy. The presented approach achieves parameter reduction results that are globally comparable to state-of-the-art solutions and outperform them across a wide range of modern neural network architectures, from convolutional models to transformers. The results demonstrate how flow divergence serves as an effective guiding principle for both filter-level and layer-level optimization, offering practical benefits for deployment in resource-constrained environments.
- 2 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (1) , 1-28
Open AccessIDAP++: Advancing Divergence-Aware Pruning with Joint Filter and Layer Optimization
Modern knowledge and large volumes of data are increasingly encoded within neural networks, making the task of simplifying their structures and reducing the number of parameters especially relevant, both to improve efficiency and to facilitate deployment in resource-constrained environments. This paper presents a novel approach to neural network compression that addresses redundancy at both the filter and architectural levels through a unified framework grounded in information flow analysis. Building upon the concept of tensor flow divergence, which quantifies how information transforms across network layers, we develop a two-stage optimization process. The first stage employs iterative divergence-aware pruning to identify and remove redundant filters while preserving critical information pathways. The second stage extends this principle to higher-level architecture optimization by analyzing layer-wise contributions to information propagation and selectively eliminating entire layers that demonstrate minimal impact on network performance. The proposed method naturally adapts to diverse architectures, including convolutional networks, transformers, and hybrid designs, providing a consistent metric for comparing the structural importance across different layer types. Experimental validation across multiple modern architectures and datasets reveals that this combined approach achieves substantial model compression while maintaining competitive accuracy. The presented approach achieves parameter reduction results that are globally comparable to state-of-the-art solutions and outperform them across a wide range of modern neural network architectures, from convolutional models to transformers. The results demonstrate how flow divergence serves as an effective guiding principle for both filter-level and layer-level optimization, offering practical benefits for deployment in resource-constrained environments.