UAV individual recognition method based on Scale Down ResNet

DOI:	10.11809/bqzbgcxb2023.06.034
Keywords:	UAV signal; individual recognition; lightweight neural network; MobileNetv2; residual network
Abstract:	In UAV individual recognition, this paper proposes a UAV individual recognition method based on Scale Down ResNet (SDRNet) to solve the problems of low classification accuracy, poor real time performance, a large number of network model parameters and difficulty in application to resource constrained equipment of the existing methods. Firstly, a one dimensional time domain signal is converted into a two dimensional time frequency image by Short Time Fourier Transform (STFT), and the image is grayed. Secondly, based on MobileNetv2 and combined with the time frequency image features of individual UAV signals, the network reduces the depth of the model by reducing the number of network layers, and reduces the dimension of the model by reducing the number of output channels. By referring to the design idea of ResNet residual structure, the lightweight residual network SDRNet model is designed by adding residual connections of the convolutional layer to achieve more integration of networks with different depths. Finally, the SDRNet model is trained by using an STFT time frequency grayscale image as a sample so as to realize the identification of individual UAVs. The simulation experiment is carried out in an AWGN White Gaussian noise channel by using the signals of 6 public hovering UAVs. The experimental results show that the average recognition rate of individual UAV signals by the proposed method is 94.00% in the SNR=10 dB environment, which is higher than 0.17% and 5.17% for MobileNetv2 and GoogleNet models respectively, and lower than 2.50% for ResNet model. The number of learning parameters, model size and multiplicative computation of the designed SDRNet model are about 19.5%, 19.6% and 35.9% of those of the basic lightweight MobileNetv2 model. Compared with the neural network models based on algorithms like MobileNetv2, GoogleNet and ResNet, the proposed method has a faster recognition speed and a lower memory overhead while maintaining a higher recognition accuracy.
Published:	2023-06-28
PDF HTML