Slimmer Deep Learning Architectures for Drone Detection

The increasing risk posed by the use of drones in illicit activities has made drone detection crucial for security and surveillance. This task, however, faces challenges such as false alerts caused by birds or distant appearances. In this context, You Only Look Once (YOLO) based algorithms, and Convolutional Neural Networks (CNNs) have demonstrated exceptional effectiveness in detecting and classifying drones and birds. Despite their effectiveness, deploying CNN models on edge devices is challenging due to memory limitations and resource-intensive demands, necessitating compression approaches like knowledge distillation and pruning. Pruning reduces memory storage requirements by eliminating unnecessary weights from the model, decreasing the model size and making it suitable for resource-constrained environments. However, this often leads to reduced accuracy. To mitigate this loss, knowledge distillation is employed. Various CNN architectures, including ResNet50, ResNet101,ResNet50v2, DenseNet121, Inceptionv3, and VGG16 are utilized with pruning and knowledge distillation. A custom dataset with 1000 drone images and 1000 bird images is considered for evaluation and the results indicate that slimmer versions of DenseNet121 and ResNet50v2, after pruning and knowledge distillation, achieved higher accuracy of 99% and 99.25%, respectively with inference speeds of 0.93s and 0.89s, respectively, on a Tesla T4 GPU. These results highlight the potential of deploying high-performance, compact CNNs on edge devices for real-time drone detection.