MobVGG: Ensemble technique for birds and drones prediction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Automated aerial activity detection, including birds and drones, is crucial for ecological surveys and collision avoidance systems.
• Existing convolutional neural networks (CNNs) often struggle with multi-class classification accuracy, particularly for distinguishing between similar aerial objects.
• Prior research has primarily focused on single-class drone detection, leaving a gap in robust multi-class identification.

Purpose of the Study:

• To develop a highly accurate multi-class classification model for distinguishing between birds and drones in aerial imagery.
• To address limitations of traditional CNNs, such as vanishing gradients and deep architectures, by proposing a novel hybrid model.
• To provide a reliable solution for automated aerial object detection applicable to both environmental monitoring and safety systems.

Main Methods:

• A novel hybrid deep learning model, MobVGG, was developed by integrating MobileNetV2 and VGG16 architectures.
• A comprehensive dataset of 4212 images each for 'bird' and 'drone' categories was curated and prepared using stringent methodologies.
• The MobVGG model was trained and evaluated for its multi-class classification performance on aerial imagery.

Main Results:

• The proposed MobVGG model achieved a superior classification accuracy of 96% for distinguishing between bird and drone images.
• Comparative analysis indicated that MobVGG outperforms existing benchmark studies in multi-class aerial object detection.
• The model demonstrated effective handling of the complexities inherent in differentiating between biological and artificial aerial objects.

Conclusions:

• The MobVGG model offers a significant advancement in multi-class aerial object detection, achieving high accuracy for bird and drone classification.
• This hybrid approach effectively overcomes limitations of traditional CNNs, providing a more robust solution for automated detection systems.
• The findings have direct implications for enhancing automated bird surveys and improving radar-based collision detection systems.