Research On Deep Learning Methods For Radar Dim Target Detection With Sea Clutter

The detection of dim targets under sea clutter has always been a difficult problem in the field of radar detection,tracking and identification.The subject has important applications in military and civilian applications.In military it can be used to detect periscope,long-distance surface ships,sea-skiing vehicles,etc.It also can be used for civil smuggling,illegal trade,and collision avoidance.Under high sea conditions,the backscattering electromagnetic scattering characteristics of sea clutter are strong,and the target signal in radar echo is easily concealed by sea peaks caused by sea waves and surges,which brings difficulties to detection.The traditional radar target detection method has many defects.There are many false alarms under low SNR.Because the sea clutter is complex and variable.The traditional method is prone to mismatch when fitting its distribution.In order to solve these problems of traditional detection methods,this thesis uses deep learning technology to transform the target signal detection problem in radar echo into binary classification problem of 'clutter' or 'clutter + target',and automatically extract clutter and network through the network.The target is classified in different characteristics of the transform domain.This thesis aims to study effective clutter suppression algorithm and feature extraction algorithm to reduce false alarms,improve detection speed,and improve radar detection performance for weak targets in complex sea environment.The specific content of this article is arranged as follows:The second chapter introduces the South African Council of Science and Industry(CSIR)radar sea surface small target detection data set.Then analyzes the time domain and frequency domain characteristics of the target.Followed by using the traditional twodimensional CFAR method to detect the target.The traditional two-dimensional CFAR method has difficulities to detect the target since the sea clutter is too strong.The main content includes the radar parameters of the CSIR data set,the wind speed sea state,the target characteristics,and the cell average two-dimensional constant false alarm rate(CACFAR)detection on the Range-Doppler plane.The third chapter proposes a radar weak target detection algorithm based on STFT and improved Tiny-Yolov2 network.It is used to solve the problems which can not be solved by CA-CFAR.Firstly,the time-frequency characteristics of the echoes are extracted by short-time Fourier transform(STFT)to obtain the time-frequency images.Because CNN has higher requirements on sample capacity and radar echo data is limited,and optical data is easy to obtain.So,the target detection problem under small sample is studied on optical dataset.The effectiveness of Tiny-Yolov2 network is verified under small samples.Forthermore,this chapter studied the sample capacity requirements for effective learning in the network.Then,for the radar time-frequency graph dataset,the improvement of filter design,data enhancement and non-maximum suppression is made for the Tiny-Yolov2 model.An improved detection algorithm based on Tiny-Yolov2 network is proposed.Finally,the effectiveness of the algorithm is verified on the radar time-frequency dataset.The results show that the network has better detection performance,lower false alarm probability and faster detection speed.The fourth chapter proposes a sea clutter suppression algorithm(DA-CCS)under the CNN framework.The convolutional layer,the pooling layer,and the upsampling layer,which retain the spatial position information,are used to extract features.The class activation map(CAM)can inversely map the network-learned tags to specific ones in the input image.The area is physically interpretable.This chapter analyzes the principles and algorithms of CNN,then deduces the network model and detection process of DA-CCS algorithm proposed in this paper.Then,the radar time-frequency image dataset is classified by convolutional layer,pooling layer and fully connected layer construction network.The algorithm is applied to different sea states,carrier frequencies,PRF,and further target.Finally,the effectiveness of the DA-CCS algorithm is verified on the measured radar dataset,to verify the effectiveness of the proposed methods.The proposed algorithm has higher detection accuracy than the recently proposed algorithms based on sparse Fourier transform.The detection speed is fast.It has the potential for real-time detection.The fifth chapter is the conclusion of the full text and the prospect of the next step.