Перегляд за Автор "Klimenko, V.I."

Зараз показуємо 1 - 3 з 3
  • Ескіз
    Документ
    Convolutional Neural Network Transfer Learning Method for Aircraft Image Classification
    (2025) Bas, I.S.; Kadynska, V.D.; Klimenko, V.I.; Mazurets, O.V.; Мазурець, Олександр Вікторович
    This study presents a transfer learning method for aircraft image classification using convolutional neural networks (CNNs), with ResNet-50 selected as the foundational architecture. In light of limited annotated datasets and diverse imaging conditions across aerial platforms, the approach leverages pre-trained models to extract generic visual features, followed by fine-tuning on domain-specific aircraft imagery. The proposed method includes preprocessing techniques, selective layer freezing, and architectural adaptation to enhance classification accuracy while maintaining computational efficiency. Experiments demonstrate improved performance over models trained from scratch, particularly in terms of convergence speed, generalisation under occlusion and variable lighting, and adaptability to various aircraft types. The solution is optimised for real-time deployment, including on resource-constrained devices. This work highlights the practical value of CNN-based transfer learning frameworks in aviation surveillance, airport automation, and security monitoring, and suggests avenues for future research involving lightweight architectures and video-based classification.
  • Ескіз
    Документ
    Neural Network Assessment of Buildings Condition Based on Visual Data
    (2025) Mazurets, O.V.; Klimenko, V.I.; Мазурець, Олександр Вікторович
    The paper addresses the problem of automated assessment of building condition based on visual data in the context of post-war reconstruction and large-scale urban monitoring. A neural network–based approach is proposed that combines automatic building segmentation in aerial and UAV imagery with subsequent classification of structural condition. Lightweight single-stage segmentation models from the YOLOv8–YOLOv12 families were investigated, with particular attention to compact configurations suitable for resource-constrained deployment. The methodology includes image tiling, data augmentation, class balancing, and a multi-head architecture for joint instance segmentation and condition classification. Experimental results demonstrate that compact models with enhanced augmentation achieve the best trade-off between accuracy and efficiency, providing reliable building localization and acceptable performance for multi-class damage assessment. The proposed approach enables automated generation of damage maps and can support prioritization of engineering inspections and decision-making in post-war reconstruction and urban infrastructure management.
  • Ескіз
    Документ
    Segmentation of textile prints with contour-stable color masks for industrial stencil printing using artificial intelligence
    (2025) Lianskorunskyi, K.O.; Klimenko, V.I.; Sobko, O.V.
    This paper presents an AI-driven approach for segmenting textile prints to obtain contour-stable binary or multi-class masks suitable for industrial stencil printing and color separation. The method combines neural network segmentation with minimalistic, topologically motivated post-processing to ensure accurate, reproducible boundaries under moderate variations in lighting, shooting conditions, and prepress transformations. Unlike traditional planar metrics, the proposed evaluation prioritizes boundary stability, contour integrity, and technological suitability for cliché preparation. Experimental results on complex textile textures demonstrate consistent mask quality, reduced manual refinement, lower ink consumption, and fewer defects in serial printing. The solution supports sustainable textile production by decreasing prepress energy and material waste while increasing process repeatability and printing accuracy. Further work will expand the dataset, refine boundary corrections for specific fabrics, and formalize end-to-end evaluation protocols from digital segmentation to control print.