High Resolution Range Profiles Recognition Based On Convolutional Neural Network

The high resolution range profile(HRRP)is the projection distribution of the complex amplitude of the target multi-scatter center in the radial direction of the radar.The HRRP reflects the projection distribution of the target cross section in the direction of the radar line of sight and the fine geometry and structural features of the target,so HRRP can be applied to the classification or recognition of the target.Target recognition based on high resolution range profile(HRRP)has become one of the most important methods in radar target recognition.Traditional target recognition faces two major problems: the cumbersome manual extraction of effective features and the difficulty of classifier design.The neural network can automatically learn the target features and reduce the complexity of the artificial design features in the traditional target recognition process.Therefore,HRRP recognition based on neural network has become one of the current research directions.In this paper,the in-depth research and simulation verification of CNN-based HRRP recognition technology are carried out from three aspects: optimization design of convolutional neural network(CNN)model,improvement of network model classification performance and enha ncement of HRRP feature extraction richness.The main contents and innovations of the paper are summarized as follows:Since HRRP recognition in the actual battlefield environment is a small sample learning problem,aiming at the problem that CNN often has a slow convergence of the loss function during training when dealing with small data sets,a convolutional neural network model suitable for small samples is proposed,and the batch normalization(BN)algorithm is introduced to optimize the performance of the model.The BN algorithm ensures the relative stability of each layer of input data and improves the learning speed of the model.To further improve the classification performance of the model,replace the Softmax classifier of the CNN output layer with a support vector machine(SVM).Experimental results using high-fidelity electromagnetic simulation data validate the effectiveness of the proposed optimized CNN model,and it is shown that when the complex model processes small samples,using SVM as the CNN classifier is more conducive to improving the recognition accuracy.According to the existing classical network architecture,it is concluded that different scale convolution kernels can be used in the same convolutional layer to obtain data features of different scales.Therefore,an HRRP target recognition architecture based on multi-scale CNN is proposed,which enriches the diversity of extracted features in convolutional layer and further improves the recognition accuracy of CNN for HRRP.Then,the Dropout operation is introduced into the fully connected layer of the architecture for model optimization,Dropout can randomly cut off the connections between neurons and control the overfitting well.Experimental results using high-fidelity electromagnetic simulation data validate the effectiveness of the proposed methods.Finally,the comparison experiment based on time complexity indicates that multi-scale CNN exchanges time complexity for the improvement of recognition accuracy,so it is more suitable for scenes with low real-time requirements but high accuracy requirements,but for scenes with high real-time requirements,a network architecture with fixed-size convolution kernels should be used.