Research On Dynamic Target Detection Of Small-sample Airborne Radar Based On Convolutional Neural Network

Motion object detection is one of the important tasks of airborne radar signal processing,but when radar works,it is often faced with strong ground clutter,which will spread in the two-dimensional domain in space time,flooding low-speed and weak moving targets.Traditional space-time adaptive processing methods can suppress the clutter of space-time coupling,but require uniform training samples as support.The actual clutter environment faced by airborne radar is often uneven,which brings difficulties to the traditional motion target detection of airborne radar.In the field of deep learning,the dynamic target classification mainly depends on convolutional neural network(Convolutional Neural Network,CNN)to data deep feature extraction,its powerful generalization ability to a certain extent weaken the influence of heterogeneous hetero on target detection,but convolutional network training needs data set and sample number is very much,it is difficult to apply to the actual airborne radar system.Based on the above problems this paper puts forward a small sample network model based on characteristic extraction network,through simulation experiments to prove that the network can well avoid nonuniform hetero for moving target detection,and in the case of low letter hybrid ratio,the network model can also well extract the main features of the moving target,effectively inhibit the clutter.This paper focuses on the above content,and the main work and contributions are as follows:(1)In the actual airborne radar echo clutter is generally heterogeneous clutter,and due to the system hardware error and the internal motion of the clutter,the received clutter signal airspace and time domain error,through the airborne radar prior knowledge and the possible error analysis,get high precision radar echo signal through high fidelity simulation.In order to better extract the echo signal features in the CNN,the echo signal is transformed to the space-time 2 D beam domain when constructing the CNN data set,and the real and virtual parts of the echo signal are jointly processed to fully extract the target and clutter features.(2)In the case of low SCR,the moving target in the radar echo is likely to be submerged in the clutter,and at the same time,there is the possibility of interference signal in the actual radar echo,which has some impact on the characteristics of the moving target signal extracted by the CNN network.To solve these problems,this paper introduces the two-channel attention mechanism of channel and space,and proposes a two-channel input-based feature extraction network model that can obtain the key information of the moving target signal,so as to improve the prediction accuracy of the moving target signal.The feasibility and effectiveness of the feature extraction network model on the prediction task are demonstrated by comparative experiments(3)Considering that there are often errors between the clutter samples generated by prior knowledge and the measured data,such as array error,carrier yaw,etc.Therefore,the actual radar echo.In order to solve the problem,this paper,on the basis of feature extraction network,the twin network model proposed a small sample network model based on feature extraction network,in the actual radar dynamic target detection can be less measured data samples to correct the prior knowledge trained feature extraction network,so as to realize the robust and fast detection.