Видання НТУ "ХПІ"

Постійне посилання на розділhttps://repository.kpi.kharkov.ua/handle/KhPI-Press/62886

Переглянути

Результати пошуку

Зараз показуємо 1 - 4 з 4
  • Ескіз
    Документ
    The issue of training of the neural network for drone detection
    (Національний технічний університет "Харківський політехнічний інститут", 2024) Hashimov, Elshan; Khaligov, Giblali
    In the article, the issue of mistaking birds for UAVs, which is considered one of the main problems in the process of detecting unmanned aerial vehicles (UAVs) through cameras, was considered. The object of the study is a method of training an artificial intelligence model so that the detection system does not mistake drones for birds. In order to solve this problem, as an alternative to creating a new complex model, the retraining method was used by making some changes to the previously developed model. The aim of the study is to train an artificial intelligence model using certain methods to detect drones and prevent them from being mistaken for birds.
  • Ескіз
    Документ
    Performance comparison of U-Net and LinkNet with different encoders for reforestation detection
    (Національний технічний університет "Харківський політехнічний інститут", 2024) Podorozhniak, Andrii; Onishchenko, Daniil; Liubchenko, Nataliia; Grynov, Denys
    The subject of study is analysis of performance of artificial intelligence systems with different architectures for reforestation detection. The goal is to implement, train and evaluate system with different models for deforestation and reforestation detection. The tasks are to study problems and potential solutions in forestry for reforestation detection and present own solution. As part of model comparison, results are presented for different artificial neural network architectures with different encoders. For training and testing purpose custom dataset was created, which includes different areas of territory of Ukraine within different timestamps. Main research methods are literature analysis, experiment and case study. As a result of analysis of modern artificial intelligence methods, machine learning, deep learning and convolutional neural networks, high-precision algorithms U-Net and LinkNet were chosen for system implementation. Conclusions. The studied problem was stated formally and broken down in smaller steps; possible solutions were studied and proposed solution was described in details. Necessary mathematical background for analysis of the performance was provided. As part of the development, accurate deforestation/reforestation module was created. All analysis results were listed and a comparison of the studied algorithms was presented.
  • Ескіз
    Документ
    Designing and evaluating DL-model for vulnerability detection in smart contracts
    (Національний технічний університет "Харківський політехнічний інститут", 2023) Shmatko, Oleksandr; Kolomiitsev, Oleksii; Rekova, Nataliia; Kuchuk, Nina; Matvieiev, Oleksandr
    Task features. Smart-contracts are programs that are stored in a distributed registry and execute code written in them in response to transactions addressed to them. Such smart- contracts are written in the Solidity programming language, which has a specific structure and syntax. The language was developed for the Ethereum platform. Having a specific structure, such languages are prone to certain vulnerabilities, the use of which can lead to large financial losses. Task statement. In this paper, a Deep Learning (DL) model is used to detect the vulnerabilities. Using the chosen approach and a properly specified input data structure, it is possible to detect complex dependencies between various program variables that contain vulnerabilities and bugs. Research results. Using well-defined experiments, this approach was investigated to better understand the model and improve its performance. The developed model classified vulnerabilities at the string level, using the Solidity corpus of smart-contracts as input data. The application of the DL model allows vulnerabilities of varying complexity to be identified in smart-contracts. Conclusions. Thus, the pipeline developed by us can capture more internal code information than other models. Information from software tokens, although semantically incapable of capturing vulnerabilities, increases the accuracy of models. The interpretability of the model has been added through the use of the attention mechanism. Operator accounting has shown significant performance improvements.
  • Ескіз
    Документ
    Usage of Mask R-CNN for automatic license plate recognition
    (Національний технічний університет "Харківський політехнічний інститут", 2023) Podorozhniak, Andrii; Liubchenko, Nataliia; Sobol, Maksym; Onishchenko, Daniil
    The subject of study is the creation process of an artificial intelligence system for automatic license plate detection. The goal is to achieve high license plate recognition accuracy on large camera angles with character extraction. The tasks are to study existing license plate recognition technics and to create an artificial intelligence system that works on big shooting camera angles with the help of modern machine learning solution – deep learning. As part of the research, both hardware and software-based solutions were studied and developed. For testing purposes, different datasets and competing systems were used. Main research methods are experiment, literature analysis and case study for hardware systems As a result of analysis of modern methods, Mask R-CNN algorithm was chosen due to high accuracy. Conclusions. Problem statement was declared; solution methods were listed and characterized; main algorithm was chosen and mathematical background was presented. As part of the development procedure, accurate automatic license plate system was presented and implemented in different hardware environments. Comparison of the network with existing competitive systems was made.Different object detection characteristics, such as Recall, Precision and F1-Score, were calculated. The acquired results show that developed system on Mask R-CNN algorithm process images with high accuracy on large camera shooting angles.