Scholarlink

<p>In aerial surveillance imagery, distinguishing drones from birds is essential to preserving airspace safety and avoiding false alarms in security systems. Despite their success in general object recognition, machine learning-based object detection techniques like YOLO and GoogLeNet frequently perform worse in drone–bird scenarios due to small objects, partial occlusion, and background clutter. To address this gap, this paper proposes a model for detecting and classifying drones (UVA) using a deep learning model. The first step involves resizing and denoising input images to ensure constant resolution and fewer noise artifacts. This is followed by using two parallel convolutional neural network (CNN) branches. Multiple convolutional, max pooling, and normalization layers are present in each branch for hierarchical feature learning. Three datasets that are publicly available with different drone types in light, dark, cloudy, sunny and desert, forest, and plains conditions are used to test the model. Compared to previous related works (Samadzadegan et al. in Aerospace 9(1):31, 2022), our model achieves enhanced performance with 92.66% accuracy, 95.4% AUC, 92.54% F1 score, 93.11% precision, and 92.78% recall at a batch size of 64.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detection and classification of UVA using double-way CNN model

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Ola M. Elsaidy,
  • Islam A. Moneim,
  • Eman I. Abd El-Latif

摘要

In aerial surveillance imagery, distinguishing drones from birds is essential to preserving airspace safety and avoiding false alarms in security systems. Despite their success in general object recognition, machine learning-based object detection techniques like YOLO and GoogLeNet frequently perform worse in drone–bird scenarios due to small objects, partial occlusion, and background clutter. To address this gap, this paper proposes a model for detecting and classifying drones (UVA) using a deep learning model. The first step involves resizing and denoising input images to ensure constant resolution and fewer noise artifacts. This is followed by using two parallel convolutional neural network (CNN) branches. Multiple convolutional, max pooling, and normalization layers are present in each branch for hierarchical feature learning. Three datasets that are publicly available with different drone types in light, dark, cloudy, sunny and desert, forest, and plains conditions are used to test the model. Compared to previous related works (Samadzadegan et al. in Aerospace 9(1):31, 2022), our model achieves enhanced performance with 92.66% accuracy, 95.4% AUC, 92.54% F1 score, 93.11% precision, and 92.78% recall at a batch size of 64.