Research On High-Resolution Radar Target Recognition Via Deep Learning Models

With the rapid development of radar technology,the high-resolution advantages brought by broadband radar are becoming more and more obvious.High-resolution radar data(such as one-dimensional high-resolution range profile(HRRP)and Synthetic Aperture Radar(SAR)images)provide richer detailed features of the target,which further provides the possibility for target's model recognition.Therefore,Radar Automatic Target Recognition(RATR)is increasingly focusing on the research of high-resolution radar.Traditional RATR technology requires a very tedious manual process and the design of complex recognition algorithms.However,the recognition effect is often poor,and the generalization ability is not good,too.With the development of Artificial Intelligence(AI)in recent years,deep learning methods have been more and more widely used in various scenarios.Convolutional neural networks have achieved great success in optical image recognition and classification tasks.In this paper,by analyzing the characteristics of high-resolution radar data,a new Convolutional Neural Network(CNN)structure is designed,and the theoretical basis of the new network,the lightweight-network and algorithm optimization schemes are given.The results had been proved by experiments.The research work of this paper is as follows:1.Inception-Based VGG Networks(IVGGNets)are proposed for high-resolution radar target recognition.This method uses VGGNets as the backbone and introduces the Inception module for the sparse characteristics of radar data.The deep point convolutional layer is also introduced into the network to enhance the ability of the method to fit nonlinear features.2.Aiming at the sparse characteristics of radar data and the reality that it is difficult to obtain large-scale radar data sets for training,Multi-Dimensional Sparse feature extraction module(MDS)is proposed.Based on MDS module,four depths of Multi-Dimensional Radar target recognition network(MDR-Net)are proposed.The MDS module introduces a larger convolution filter.Compared with IVGGNets,MDR-Nets can achieve higher accuracy,but greatly increases the amount of network parameters and calculations.3.In response to the problem of excessive computing cost on MDR-Net,Lightweight Multi-Dimensional Sparse feature extraction module,(LMDS)module is proposed.The LMDS module adds shallow point convolution layer follow the large convolution filter in MDS module.It reduces the number of feature maps,thereby greatly reducing the amount of floating point calculations required in the module.This paper also optimizes the classifier,reduces the number of parameters through utilizing one fully connected layer and global average pooling method,and optimizes network performance by introducing hyperbolic tangent function and gradient centering method.To verify the effectiveness of these methods,large-scale experiments have been carried out,some of which are controlled experiments with controlled variables.Finally proved that IVGGNets has better performance than VGGNets and other networks.The performance of MDR-Nets is better than IVGGNets,but with higher computational cost.The LMDS module can realize the lightweight optimization of the MDS module,but the lightweight network will have a slight performance loss and make the training process more tortuous.The training picture can be smoothed by introducing a gradient centering algorithm and the network can converge in advance.