| Radar target detection has always been a focal research area in radar signal processing.Traditional detection algorithms are generally based on statistical detection theory,which views clutter as a random process and estimates its distribution model.However,when the estimated clutter model does not match the actual clutter model,it makes it difficult for classical constant false alarm rate(CFAR)detectors to detect the presence of targets.In recent years,deep learning technology has developed rapidly,and many researchers have applied deep learning technology to radar target detection.This thesis focuses on the problem of target detection in sea clutter background and study the clutter suppression method and target detection method based on neural networks.The specific work contents are as follows:1.We investigated clutter suppression methods based on fractional-order singular value decomposition and the classical CFAR detection algorithm for actual sea clutter data.Experimental results show that the clutter suppression method based on fractional-order singular value decomposition has a certain suppression effect under different signal-toclutter ratio(SCR)conditions,while the traditional CFAR detection algorithm requires a reasonable false alarm rate to achieve better detection performance.2.To address the problem of high clutter power in complex sea surface environments,which affects target detection,a clutter suppression method based on residual generative adversarial network(GAN)networks was proposed.The deep convolution generative adversarial network(DCGAN)networks is used to learn and generate simulated sea clutter signals from actual sea clutter,which solves the problem of a small number of samples in the sea clutter dataset.By alternately training the suppression network and the discrimination network,the weights of the suppression network are obtained under different SCR conditions,and then performed clutter suppression on the signal.Experimental results show that compared with traditional clutter suppression methods,as well as clutter suppression methods based on deep convolutional neural network(DCNN)and DCGAN,the proposed method has better clutter suppression performance under low SCR conditions.3.To address the problem of decreased detection performance of the classical CFAR detection algorithm under strong clutter background and the long computation time of existing deep learning-based radar target detection methods,a two-step detection method based on deep neural networks was proposed.Firstly,small-size samples are obtained by sliding window segmentation of the range-doppler map of the radar echo,and preliminary detection is performed through the fully connected layer network(FCNN).Then,largersize samples are obtained from the detection area for deep feature extraction,and the target and clutter are distinguished through a multi-layer convolutional neural network(CNN)to obtain the final detection results.The experimental results show that compared with the classical CFAR detector and the one-step detection method based on CNN,the proposed method maintains a high detection probability while effectively improving the detection efficiency. |
