Research On Variant Target Recognition Based On Deep Learning

The development trend of modern war is information and intelligence.As an important equipment in modern warfare,the role of radar is not only to provide the spatial position of the target,but also to obtain more detailed information of the target through the echo signal.Radar automatic target recognition technology can obtain information such as the attribute,category and even model of the target.High resolution range profile(HRRP)contains structure information such as target size and scatterer distribution,which is easy to obtain and process,so it is widely used in radar automatic target recognition.The high resolution range profile of the target will change with the change of its shape.On the battlefield,due to the different combat missions,the configuration of the target will be adjusted,and its appearance will change accordingly.There is a problem of structural mismatch between the deformed target and the high-resolution range profile of the same model target in the recognition database,which will lead to a significant decrease in recognition performance.Aiming at the problem of target recognition of this kind of variant,this thesis has carried out research,and the main work contents are summarized as follows:(1)The HRRP before and after the deformation of the target is analyzed by electromagnetic calculation simulation,and it is found that the HRRP before and after the deformation of the target still has the similarity,and the variant component that changes has the structural sparse property.Then the variant target signal is modeled and regarded as the sum of the original target signal and variant component.Therefore,two ideas are proposed to realize robust identification of variant targets.One is to estimate and remove the variant component in the variant target echo and reconstruct the undeformed target echo for recognition.Another is to design a robust classifier of variant component for end-to-end identification.(2)Two recognition methods are proposed to solve the problem of recovering the original target signal by removing variant components.The first is a variant target reconstruction recognition method based on long short-term memory(LSTM)network.In this method,the original target HRRP is represented by the sparse coefficient of a specific dictionary matrix,and the variant component is modeled by LSTM network.In order to solve the model,an alternate iteration algorithm is proposed.Firstly,the variant component is fixed and the sparse representation of HRRP of the original target is estimated by using the orthogonal matching pursuit algorithm.Secondly,HRRP of the original target is fixed,the correlation of non-zero elements of variant component is captured by LSTM,and the variant component is estimated by the least square method.After the algorithm converges,the original target HRRP is used for identification.The second is to use the HRRP before and after the deformation of the target as the training sample to train the depth network.By using the strong nonlinear mapping ability of the depth network,the variant component is removed and the original target HRRP is recovered for identification.Simulation results show that the two methods can reconstruct the original target signal accurately and improve the recognition performance of the variant target.(3)Robust classifier is designed for end-to-end recognition.Using the sample joint training depth network before and after the target deformation,the robust classification features of variant components are extracted from the samples,and then the variant targets are identified end-to-end.The experimental results show that the spatial distribution of the features proposed by the network changes little with the deformation of the target,and it has good robustness.The average correct recognition rate of the end-to-end recognition of the variant target using the depth network is further improved compared with other reconstruction recognition methods.The method proposed in this thesis can effectively reduce the influence of variant components on the recognition performance,solve the problem that the recognition performance of the variant target is obviously reduced due to the structural mismatch between the variant target and the target of the same type in the database,and effectively improve the recognition performance of the variant target.