Study On Synthetic Aperture Radar Target Robust Recognition Based On The Design Of Deep Networks

As an active microwave detection radar,Synthetic Aperture Radar(SAR)realizes highresolution imaging through virtual aperture synthesis,breaks through the limitation of a single aperture of radar,has the advantages of all-weather,all-weather,long-distance action,and can provides rich information about the shape,structure,and scattering characteristics of the target.In order to interpret SAR images,SAR Automatic Target Recognition(ATR)came into being and plays an important role in typical military applications such as target reconnaissance and battlefield situational awareness.Traditional SAR target recognition methods generally extract manual features according to the characteristics of SAR images,and then input them into the classifier for recognition.The manual feature extraction process relies heavily on expert knowledge and the process is cumbersome,which makes it difficult to meet the application requirements of automation and intelligence.Since deep learning can automatically extract features from data and has a strong feature fitting ability,it has been widely introduced into the research of SAR target recognition.Although deep learning has powerful feature extraction capabilities,most of the current SAR target recognition methods based on deep learning are not targeted at low signal to clutter ratio(Signal to Clutter Ratio,SCR),variant expansion,and limited annotations.Facing the actual situation,the recognition performance drops significantly,which cannot meet the requirements of robust recognition.Aiming at the problem of insufficient robustness in the three cases of low SCRs,variant expansion,and limited annotations when deep learning methods are applied to SAR target recognition,this paper achieves robust recognition in three aspects by designing effective deep networks.The main content of this article is summarized as follows:1.Research on the problem of insufficient recognition robustness of deep learning methods applied to SAR target recognition under low SCRs.The deep network generally extracts features from the input data as a whole without considering the influence of clutter.This will result in redundant features extracted from clutter that will affect the recognition,making the recognition robustness of the model insufficient when the SCRs are low.To address this problem,the third chapter of this paper designs a SAR target clutter robust recognition method based on Point-wise Gated Convolutional Auto-encoder(PGCAE).The proposed PGCAE model includes target area extraction module,label learning module and reconstruction learning module.The target area extraction module uses the convolutional autoencoder model to learn selection factors to distinguish between target and clutter areas;the label learning module further extracts features from the extracted target area and maps them to labels for learning,helping the model extract separable features for recognition;The structure learning module performs reconstruction learning on the target area and the clutter area at the same time.The proposed model is optimized using an end-to-end learning approach.Experiments based on the recognition under different SCRs of the MSTAR measured data set show that the proposed method has better recognition performance under low SCRs,and can more accurately extract the target area.2.Research on the problem of poor recognition robustness under variant expansion conditions when deep learning methods are applied to SAR target recognition.In the practical SAR target recognition environment,it is impossible to guarantee that the target to be recognized is completely consistent with the target collected in the training phase in structure configuration,and these targets with structural differences are called variant targets.The current SAR target recognition method based on deep learning usually directly uses the optical model to recognize SAR images,and lacks the mining of the physical structure characteristics of the target,so the performance of variant target recognition is poor.The electromagnetic scattering model of the SAR target describes the physical characteristics of the target from the imaging point of view,and the scattering center features extracted based on the electromagnetic scattering model can well reflect the physical structure information of the target,which has advantages for variant target recognition.However,the existing recognition methods which use scattering center features are insufficient for mining target structural features,resulting in limited performance in recognizing variant targets.To address the above problems,the fourth chapter of this paper considers that the distribution of scattering centers can well reflect the structural characteristics of the target,and the structural characteristics of the target can be modeled through the graph convolutional network.Therefore,the point scattering center model,attribute scattering center(ASC)are respectively combined with a Graph Convolutional Network(GCN)for robust identification of SAR target variants.1)The first proposed method first extracts the point scattering center from the SAR target,and models the point scattering center as graph data,of which the scattering center parameters are adopted as the node features.Then a multi-scale GCN combined with label smoothing is utilized to extract the structural features of the target from the constructed graph data,which enables robust identification of variant targets.2)The second proposed method proposed considers that although the first method utilizes the scattering structure features of the SAR target,it does not make full use of the visual features,so the second method proposes to combine the visual features and scattering structure features of the SAR target for recognition,and further extends from point scattering center modeling to attribute scattering center modeling.The proposed method consists of three parts: visual feature extraction module,scattering structure feature extraction module and feature fusion module.The visual feature extraction module uses a convolutional neural network(CNN)to extract visual features from SAR images.The scattering structure feature extraction module first extracts the ASCs from the SAR image,and models each ASC as a node to construct the graph data.Considering that the ASC parameters do not fully reflect the local features,the visual features corresponding to each ASC in the shallow CNN feature map are multiplied point by point by the ASC reconstruction map as the node features of the graph data.Scattering structure features are then extracted from the constructed map data by multi-scale GCN,and further fused with global visual features for SAR target recognition.Experiments based on the measured MSTAR dataset show that,compared with other SAR target recognition methods using scattering center features and traditional SAR target recognition methods,the proposed two methods have better recognition performance under the condition of variant expansion.3.Research on the problem that the existing deep learning method relies on a large amount of labeled data when it is applied to SAR target recognition,and the recognition robustness is poor in the case of a small number of labeled samples.Existing semi-supervised SAR target recognition methods generally select high-confidence unlabeled samples for pseudolabel learning,resulting in a large number of low-confidence samples that are not fully utilized.In response to the above problems,the fifth chapter of this paper proposes a semisupervised SAR target recognition method that combines complementary label learning and contrastive learning.First,low-confidence samples are given complementary labels based on their minimum predicted probabilities for utilization based on complementary label learning.By setting the threshold more accurate complementary labels can be selected to alleviate the impact of complementary label erroneous samples on model learning;in addition,the proposed method designs a contrastive loss combined with complementary labels based on complementary label learning,which makes full use of the characteristics of complementary labels compared to traditional contrastive losses.Richer negative sample pairs can make full use of low-confidence samples,which can further improve the recognition performance.Experiments on the various number of labeled samples based on the measured MSTAR dataset show that,compared with other semi-supervised recognition methods based on deep learning,the proposed method has better recognition performance under the condition of a small number of labeled samples.