Research On Open Set Recognition Of Radar High Resolution Range Profile

Radar automatic target recognition(RATR)based on high resolution range profile(HRRP)has become an indispensable part of modern radar technology due to its important practical application value.If our radar can timely identify the classes and types of enemy aircraft,missiles,ships and other targets through RATR technology,it is of great significance for our commanders to make timely high-quality responses and occupy the battlefield initiative.At present,the task of RATR mainly focuses on closed set recognition,which defaults that the targets captured by the radar belong to the known classes of the database.However,it is very difficult to obtain an HRRP database that contains a large enough number of targets,especially for non-cooperative targets.Therefore,in practical applications,radar will capture the unknown targets that have not been trained in the training phase.If closed set recognition is directly applied to RATR,it will be difficult to effectively detect the unknown targets,and will misjudge the unknown targets as a certain known class with high confidence.Such misjudgments will cause serious information interference to our commanders on the battlefield,thus causing unnecessary losses.Therefore,RATR requires that the recognition model can not only recognize the known targets,but also detect the unknown targets,which is the task of open set recognition(OSR).In short,without completely solving the technical problems of OSR,RATR can only stay in the paper stage,which will greatly affect the process of our national defense intelligent information.In this paper,the OSR method based on deep learning technology is studied,and the basic principle of using neural network to address the OSR problem is specifically studied,which concludes 2 important technical routes for improving the OSR performance: designing feature extractor which can avoid overlapping of the known and unknown features,and designing a classifier that can determine the closed classification boundary of the known features.This paper proposes 5 feature extractors and3 closed classification boundaries.In addition to the prototype boundaries inherent in prototype learning,this paper provides 20 OSR performance measurements under5 feature extractors and 4 closed classification boundaries,which greatly advances the work progress in the field of OSR based on HRRP.The first part of this paper studies the basic principle of using neural network to address the OSR problem,which finds that the unknown features always tend to be distributed in the central region of the feature space.This law persists even after changing data types,neural network structures,and loss functions.In this paper,the matching theory of neural network is proposed to explain this law,which regards the complex calculation process of neural network as a matching system.The training of neural network is the process of promoting the matching between the training sample and the network.The mismatch between the unknown targets and the trained network will cause internal friction of the system,which will reduce the output of the network,resulting in the distribution rule of the unknown features.Therefore,if the network can control the known features in the edge region of the feature space,and the unknown features will fall in the central region,then the overlapping degree of their features can be effectively reduced,and the known and unknown classes can be effectively distinguished.Under the guidance of the exploration results of OSR principle in the first part,3OSR methods based on constrained feature prototype learning are proposed in the second part of this paper,namely spatial location constraint prototype loss(SLPCL),surrounding prototype loss(SPL),and kinetic prototype framework(KPF).Specifically,SLCPL and SPL add a feature distribution range constraint term to compress the distribution range of the known features on the basis of the prototype learning classification loss term,so as to reduce the overlap probability of the known features with the unknown features.To further improve the OSR performance,SLCPL and SPL control the distance from the prototypes to the space center to make the prototypes to be distributed the surrounding region of the feature space.By introducing a new distance setting,KPF makes the spatial origin,prototype center and sample features appear in the shape of three points and one line,and directly constrains the known features in the surrounding region of the feature space.The boundary constraint radius contained in the model can change with the optimization of the model,which can be understood as the motion mode of the model,which is called kinetic pattern in this paper.In the third part of this paper,the adversarial kinetic prototype framework(AKPF)and its updated version AKPF ++ are proposed by combining the adversarial learning strategy with the proposed method KPF.AKPF model can generate adversarial samples and add these samples to the training phase,which can improve the model performance along with the adversarial motion of the boundary constraint radius.Compared with AKPF,AKPF++ further improves OSR performance by adding more generated data in the training phase,introducing more complex adversarial learning strategies and kinetic patterns.The fourth part of this paper studies the OSR method based on closed classification boundary,and 3 different methods are proposed,which are hypersphere boundary,extreme value boundary and detection boundary.Based on the prototype boundary,the center of the hypersphere boundary is shifted from the prototype to the feature center of the known feature cluster,which improves the unreasonable location of the boundary center.However,this boundary defaults that the rate of decline from the center to the edge of the class probability distribution is determined by the negative e exponential.This empirical model lacks theoretical basis.In this paper,the extreme value boundary theorem is proposed and proved,and the extreme value boundary is proposed.It is proved that the rate of decline from the center to the edge of a given class probability distribution should be determined by the cumulative distribution function of the generalized extreme value distribution.The above model describes cluster boundaries in a way that is not precise enough to be suitable for high-dimensional feature clusters with complex boundaries.Therefore,this paper proposes a detection boundary algorithm that can detect high-dimensional feature cluster boundary samples pointto-point,which determines a compact boundary around the cluster according to the distribution balance and distribution density of the nearest neighbor objects of the point to be detected.