Mid-course Target Recognition And Geometric Feature Inversion Based On Deep Learning

Ballistic missile will go through three stages after launch: boost stage,mid-course and reentry stage.The mid-course takes the longest time and farthest from the ground,which is the best opportunity to detect and identify ballistic missiles.The study of midcourse target recognition and geometric parameter inversion is the most important part of ballistic missile defense system research.In this paper,the convolution neural network(CNN)and long short-term memory neural network(LSTM)are designed to identify the target for the two-dimensional ISAR image of the middle target and the one-dimensional time-domain electric field echo.After the successful recognition of the target,this paper also studies the geometric characteristics inversion method of the middle target based on CNN.The main research contents are as follows:(1)A convolution neural network(CNN)based recognition algorithm to identify ISAR image of mid-course target is proposed.This method first simulates the electromagnetic scattering of targets and acquired the scattering electric field of targets at different frequencies and scattering angles,and constructs the target electromagnetic scattering model to obtain the ISAR image of the target.Then the target ISAR image is input as a data set to the CNN model for training.The final result shows that the accuracy of the mid-course target recognition based on the CNN model is 97.29%,which is much higher than the 76.11% of the SVM.In order to use the information between the polarizations of the ISAR image,it is proposed to combine the ISAR images of the target under different polarizations into a sample and input it to CNN for training.CNN is used to automatically extract the information related to the polarization.The target recognition accuracy rate is 99.72%,which verifies the reliability of the method.(2)A LSTM model is proposed to extract the features of the target electric field echo and identify them,and the attention mechanism is used to improve the influence of key features in the model on the final recognition results.The final recognition accuracy is 98.88 %.In order to optimize the model,the time-domain electric field echo framing method is proposed to change the number of parameters input to each LSTM unit so as to reduce the parameters in the model without reducing the amount of sample data and improve the training efficiency.The final results show that the framing method can improve the accuracy of target recognition while reducing the model complexity.(3)A method of geometric feature inversion of mid-course target based on convolutional neural network is proposed.This method takes the ISAR image of the target as a sample,and designs a CNN model to extract features and predict the height and sectional area of the target.Compared with traditional methods,using one CNN model can simultaneously invert multiple geometric features,which greatly improves the inversion efficiency.Finally,the correlation coefficient between the predicted height and the true height is 0.9907 and the correlation coefficient between the predicted value of the target sectional area and the true value is 0.9754.The experimental results show that this method can effectively predict the geometric characteristics of the target,and provide a new idea for the inversion of the geometric features of the midcourse target.