| Radar high resolution range profile(HRRP)is the coherent summations of the complex returned echoes from target scatters in each range resolution cell projected on radar line of sight after wideband radar signal pulse compression.It contains much important information,such as target size,structure,scattering distribution,etc.HRRP has advantages of easy acquisition and few storage consumption.Hence,HRRP target recognition has become a hotspot of research in radar automatic target recognition(RATR)area.How to extract effective features from HRRP data is the key problem of HRRP target recognition.In this paper,according to the characteristics of HRRP,the problem of feature extraction and target recognition methods based on HRRP are studied.The main content of this paper is summarized as follows:1.The HRRP target recognition method based on recurrent neural network(RNN)is studied.Considering the correlation between the range resolution cells of HRRP data,the preprocessed HRRP is transformed into a sequence form through sliding window processing.RNN is used to model the sequential data,to extract the structural features and internal correlation of HRRP.The hidden layer activation of the last time step is used as the final recognition feature,which pass the softmax classifier to give the category of target.Examining the recognition performance of the proposed model on measured data.The values of key parameters in the proposed model are discussed.Results show that compared with two common target recognition methods,the features extracted by RNN have stronger separability,so the recognition performance is better.2.The HRRP target recognition method based on attention mechanism is studied.Considering the differences in separability of different regions in HRRP data,the attention mechanism is adopted to make full use of each region’s information for target recognition,according to the degree of importance of different regions to recognition.Concretely,two models are proposed.The first model uses soft attention to get a weighted sum of hidden layer activation of RNN at each time step,aggregating the features of all segments in the sequential HRRP.The second model expresses the correlation and separability of each region through multiple operations of self-attention and multi-layer perceptron,making the network give more accurate recognition results according to the data segments with strong discrimination.Discussing the key parameters of the proposed model on measured data.Recognition accuracy and visualization of attentional distribution verify that attention mechanism can enhance feature extraction capability of the network,so as to achieve better recognition performance.3.The HRRP denoising and target recognition method based on generative adversarial network(GAN)is studied.Considering that the HRRP obtained in real application contains noise,it has bad effect on the accuracy of the recognition system.GAN is used to denoise the HRRP with low SNR,and the denoised HRRP is processed by the self-attention-based recognition model to obtain corresponding target recognition result.In order to reserve signal components beneficial to recognition,the proposed model combines the process of denoising and recognition by hybrid loss.Through end-to-end training,integrated HRRP denoising and target recognition framework is achieved.Three data sets with different SNR are created to examine the proposed model.The analysis of experimental results verifies that the proposed model has good denoising and recognition performance under low SNR scenarios and the hybrid loss ensures the integrity of recognition features to a certain extent. |
