Research On Radar Target Detection Technology In Complex Scenes Based On Deep Learning

As the basic task of radar signal processing,radar target detection has vital applications in both the military and civilian fields.In traditional detection methods,statistical inference based on probability distribution is used to distinguish the target of interest from clutter.But there are obvious performance bottlenecks in complex environments.Deep learning technology has good application potential in radar target detection since it can solve the problem that complex background clutter cannot be accurately modeled,such that the detection performance in complex backgrounds can be improved.However,as a data-driven adaptive technology,the radar target detection method based on deep learning is limited by the quality of the data set.Thus,there are high-performance detection problems in complex scenes and adaptation problems of network detection in changing scenes.In this thesis,the radar target detection technology in complex scenes based on deep learning is studied,which provides a theoretical method and technical way to improve the performance of network detectors in complex scenes.Based on the analysis of clutter and target properties,a constant false alarm rate detection method based on neural network ensemble and two data adaptive network detection methods in changing scenes are proposed.Specifically,the main contents are summarized as follows.1.Aiming at the high-performance detection problems in complex scenes,a constant false alarm detection method based on neural network ensemble is proposed.Taking the amplitude,the power spectral density,and the time-frequency maps as input,this method adopts the stacking strategy for sub-network generation and ensemble,and then realizes constant false alarm rate detection.Using the IPIX dataset,this method is verified to have superior detection performance in real complex scenes.2.Aiming at the adaptation problems of network detection methods in changing scenes,an adaptive network detection method combined with scene classification is proposed.In this method,the scene classification network is designed to control the output ratio of a group of detection sub-networks,so as to maximize the posterior probability of the feature vectors.Simultaneously,feature fusion and detection are realized.Using simulation and measured data,this thesis analyzes the advantages of the proposed method in data feature fusion,and verifies that the proposed method has superior detection performance in changing scenes.3.Aiming at the adaptation problems of network detection methods in changing scenes,an adaptive network detection method based on multi-task learning is proposed.This method integrates contrastive learning into a multitask auto-encoder.It can learn a compact and distinguishable feature representation,which enables similar data to be clustered and different data to be distinguished.Simultaneously,a classifier is introduced to perform a binary detection in the feature representation.Using simulation and measured data,this thesis analyzes the advantages of the proposed method in data feature representation and enhancement,and the adaptability performance in changing scenes is verified.In conclusion,this thesis enriches the radar target detection method based on deep learning,provides ideas for the extraction and representation of the distinguishability between target and clutter,and provides a theoretical method and technical implementation for breaking through the performance bottleneck of target detection in complex backgrounds.