Research On Classification Method Of Narrow-band Radar Target Based On Deep Feature Fusion Network

Modern narrow-band radars can acquire the micro-Doppler signature caused by the target's micro-motion,which is regarded as the unique the characteristics of the target.Therefore,technology of narrow-band radar target classification has achieved unprecedented development.The traditional narrow-band radar target classification methods extract the typical features of the target echo in a single feature domain such as time domain,Doppler domain and so on,which is then classified by a classifier.However,features from a single feature domain can't reflect all the characteristics of the target,and have a dependency on the quality of the feature domain,which will eventually affect the classification performance.Note that the multiple feature from different feature domains can not only reflect the target characteristics more comprehensively,but also enhance the robustness of the classification.However,fusing different features from multiple feature domains is a complex problem,which cannot be solved by the traditional fusion methods.Therefore,this thesis focuses on the researches of feature fusion from multi-domain features.The main work of this thesis is as follows: 1?Aiming at the multiple domain feature fusion problem in narrow-band radar target classification,we propose a hierarchical fusion network(HFN)in this thesis.The HFN consists of two parts: an intra-domain fusion network and an inter-domain fusion network.The intra-domain network composed of auto-encoder reduces the redundancy of the features in each feature domain on the premise of the separable information being reserved.Moreover,in order to increase the separability of fusion features,the constraint of the ratio of within-class distance and between-class distance is introduced.In particular,the inter-domain network fuses the features from different feature domains through a multi-layer perceptron,and then a classification result is obtained.In addition,the intra-domain network and inter-domain network are optimized jointly,in which the feature fusion module and the classifier are combined with each other.The recognition rate of the HFN on the two measured narrow-band radar datasets,including the ground target dataset and aerial target dataset,are 90.28% and 96.41% respectively,which exceeds the existing narrow-band methods and verifies the effectiveness of the proposed method for the target classification performance.2?To achieve the purpose of features fusion,traditional feature fusion methods usually adopt the way of weighted summation to discrete features,and then perform non-linear transformation.Whereas for the disperse features,the multiplication operations of second and higher order terms are rarely considered.The addition of features is more inclined to the union of features,which makes the larger feature values have a greater impact on fusion.However,the multiplication of features is more inclined to the intersection of features,which means the combination of features with large feature values has a relatively greater influence on fusion.Although the above-mentioned hierarchical fusion network has been demonstrated the promising performance on narrow-band radar target classification,there is still room for improvement.In this thesis,the combination of addition and multiplication operations is considered to improve the hierarchical fusion network.This method adds a second-order term fusion module to the inter-domain fusion network.The second-order term considers all second-order combination ways,which makes the inter-domain fusion more sufficient.Finally,based on the two measured narrow-band radar datasets,the recognition rates are 90.90% and 97.82% respectively,which validates the effectiveness of the method.