Drone Recognition Using Deep Learning Methods

Bird and drone detection is a critical visual classification task in intelligent monitoring systems, with practical applications in airport surveillance, low-altitude airspace management, and public safety. However, reliable classification remains highly challenging due to the typically small size of flying targets and their susceptibility to changes in viewpoint, motion blur, background interference, and significant intra-class variation in bird samples. In this paper, we systematically compare the performance of eight deep learning models - including four convolutional neural networks and four visual Transformer models - on the bird-drone closed-set classification task under a unified experimental protocol. All models were initialized with ImageNet pre-trained weights and trained and tested under the same data partitioning, training strategies, and evaluation metrics. The results indicate that all models achieved high classification accuracy on the current dataset, but significant differences still exist among different architectures. The optimal model achieved the highest classification accuracy, while models exhibiting a better balance between accuracy and training efficiency demonstrated superior overall performance. Confusion matrix analysis further indicates that classifying the "Bird"class is generally more challenging than the "Drone"class, with the primary error pattern being the misclassification of 'Bird' as "Drone."These results provide a clear closed-set comparison baseline for bird-drone recognition and offer guidance for model selection in small-object visual classification tasks.