Assessing and Optimizing Deep Learning Models Performance for Drone Detection

The extensive spread of unauthorized drones exposes formidable security threats, requiring effective and efficient drone detection and identification systems. Existing drone detection systems have lower accuracy in distinguishing between visually similar categories such as birds and drones. This research presents deep learning-based approaches for drone detection using EfficientNet-B0, DenseNet-121, and Visual Geometry Group-16(VGG-16). This research aims to improve classification and accuracy performance. These models use an aerial image dataset including drones and non-drones and employ transfer learning to enhance feature extraction and classification accuracy. These models are validated via K-fold cross-validation and experimental results prove EfficientNet-B0 DenseNet-121 and VGG-16 perform better compared to existing models’ accuracy, precision, recall, and F1-score. EfficientNetB0 reported a maximum accuracy of 97%, showing strong performance compared to VGG-16 (95.9%) and DenseNet-121 (94.04%). Each model demonstrated strengths across different performance metrics.