Few-Shot Signal Recognition for UAV Based on Prototype Network

Abstract

In recent years, the frequent occurrence of illegal use of unmanned aerial vehicles (UAVs) has posed a serious threat to societal security. Therefore, the identification and classification of UAVs have become critically important. However, UAV signal recognition based on deep learning has faced technical challenges due to the lack of large-scale UAV datasets. To address this issue, a prototype network based on a composite loss is proposed to tackle insufficient samples in UAV signal recognition. The proposed network combines cross-entropy loss and label smoothing loss, and then utilizes a Residual Network (ResNet) for feature extraction of UAV signals. To evaluate the performance of the proposed network, we compare it with the prototype network only with cross-entropy loss, and conduct tests with 1, 5, 10, 15 and 20 samples. Experimental results demonstrate that the proposed network exhibits excellent performance in dealing with few-shot classification problems. Particularly, when the number of samples exceeds 15, the accuracy of the proposed network can reach 98%.