High Resolution Range Profile Target Recognition Based On Complex Deep Network

The high resolution range profile(HRRP)contains structural information such as the target’s scattering point distribution and radial size,and has the advantages of being easy to obtain and easy to handle.Therefore,HRRP target recognition has become one of the important development directions in the radar field.In recent years,deep learning has flourished.Considering that deep networks can extract deep abstract features and have obvious advantages over traditional methods in generalization and robustness,the radar field has also tried to apply deep networks to HRRP target recognition.Existing deep network recognition methods mostly perform feature extraction in the real number domain.In this process,the phase information is discarded,which has a certain impact on the final recognition effect.For this reason,this paper introduces a complex deep network to recognize complex HRRP data to improve the recognition performance.The main content of the full text is as follows:1.In terms of modeling the global structure of complex HRRP,the densely connected convolutional network(DenseNet)is extended to the complex domain,and a HRRP target recognition method based on the complex DenseNet model is proposed.On the basis of the real number DenseNet network architecture,the complex number activation function,complex number convolution and complex number batch normalization are used to replace the original real number operation,while retaining the dense connection structure,constructing the complex number DenseNet network.Based on the measured data,the recognition experiments were carried out using the one-dimensional complex HRRP feature and the two-dimensional time-frequency spectrum feature respectively.Compared with real number convolutional neural networks,real number DenseNet,complex number convolutional neural networks(CNN)and other networks,the results show that complex number DenseNet has better recognition performance under different signal-to-noise ratios and different numbers of training samples.2.In the complex number HRRP sequence structure modeling,the real number gated unit(GRU)is extended to the complex number domain,and a complex number gated RNN based HRRP target recognition method is used.Based on the structure of the real GRU,the network uses complex activation functions to replace the real activation functions in the reset gate and the update gate,forming a complex gated RNN.Based on the measured data,the recognition experiments were carried out using the one-dimensional complex HRRP feature and the two-dimensional time-frequency spectrum feature respectively.The results show that the average correct recognition rate of the complex-gated RNN is significantly improved compared to the real-number recurrent neural network,real-number GRU,and real-number long short-term memory(LSTM).