Факультет математики, фізики та інформаційних технологій

Постійне посилання на фонд

https://dspace.onu.edu.ua/handle/123456789/37442

Переглянути

Перегляд Факультет математики, фізики та інформаційних технологій за Автор "Feng Xuejiao"

Зараз показуємо 1 - 1 з 1
  • Ескіз
    Документ
    Analysis and Implementation of Objiect Detection Information Technology Based on Deep Learning
    (Одеський національний університет імені І. І. Мечникова, 2024) Фен Сюецзяо; Feng Xuejiao
    The topic of "Analysis and Implementation of Object Detection Information Technology Based on Deep Learning" is considered in the thesis. Deep learning-based object detection technologies have become widespread, particularly in the field of autonomous driving and intelligent transportation systems. These technologies offer significant advantages in terms of accuracy and efficiency, enabling real-time detection and recognition of various objects, including vehicles, pedestrians, and traffic signs. However, despite their advantages, such systems face challenges in complex environments, such as foggy weather, low light conditions, and dense traffic scenarios, which can degrade detection performance. The purpose of the work is to improve existing deep learning-based object detection algorithms to enhance their robustness and accuracy in challenging conditions. As a result of the research carried out in the work, various state-of-the-art object detection algorithms were analyzed, and novel methods for improving detection performance in foggy conditions and lightweight network architectures were proposed. Specifically, the integration of dehazing techniques, attention mechanisms, and optimized loss functions were explored to enhance the detection accuracy and efficiency of the YOLOv5 and YOLOv8 models. Additionally, the work included the construction of a foggy scene vehicle detection dataset and the implementation of a model system for simulating and evaluating the proposed methods. In the work, a number of requirements for the class library, which can be used in deep learning-based object detection systems, as well as for the model system, which made it possible to simulate the considered methods, were formed. Both systems were implemented. Using the proposed methods, it was possible to achieve significant improvements in object detection accuracy under foggy conditions and in lightweight network architectures. The integration of dehazing techniques and attention mechanisms allowed for better feature extraction and focus on key areas, while the optimized loss functions improved the model's ability to detect small and occluded objects. The lightweight network architectures demonstrated a good balance between computational efficiency and detection accuracy, making them suitable for deployment on resource-constrained devices. Overall, the research provides valuable insights and practical solutions for enhancing the performance of deep learning-based object detection systems in realworld applications.