Sea Clutter Suppression Based On Sparse Representation And Deep Learning

As our army moves toward deep blue and the national marine development strategy continues to advance,there is an urgent need to improve the radar's sea surface target detection capability.Sea clutter suppression is of paramount importance in the field of marine target detection.The sea clutter has complex characteristics and strong echo amplitude,which results in low signal to clutter ratio(SCR)of the sea surface weak target,which is easy to be submerged in the sea clutter background,so the target detection is difficult.The traditional sea clutter suppression approach is mainly designed based on the difference between target and sea clutter energy or Doppler spectral distribution characteristics.However marine moving target falls into the clutter Doppler channel easily,and it is difficult to separate target from sea clutter through the time domain or the frequency domain.Therefore,it is of great theoretical research significance and engineering application value to explore the new sea clutter suppression technology and realize long-range detection of radar in strong sea clutter background.In this thesis,the sparse representation theory and the deep convolutional network are used to investigate sea clutter suppression technology and marine target detection methods.The research work of the thesis is mainly divided into the following three aspects:1)In-depth analysis of the complex physical mechanism and statistical correlation characteristics of sea clutter,and based on the sparse representation and deep learning theory,the oscillation characteristics and the deep network features of the sea clutter and the target are presented.The analysis of these characteristics provides theoretical support and algorithm design reference for the subsequent sea clutter suppression algorithm.2)The sea clutter suppression technology based on sparse representation is studied.On the basis of the difference of target and sea clutter oscillation characteristics,the sparse representation of sea clutter and target can be obtained through the tunable Q-factor wavelet transform(TQWT).An energy selection method is proposed to optimize the target's wavelet representation coefficients to achieve effective reconstruction of the target.Aiming at the sea surface target in the main clutter region,a selection criterion of quality factor Q which is easy to be applied in engineering is given,and the generalized design of Q factor in the main clutter region is realized.A fractional Fourier transform domain is proposed.The selection criterion of quality factor Q is discussed for sea surface targets in the main clutter area.And the improved TQWT algorithm based on fractional Fourier transform(FRFT)domain is proposed to solve the problem of detecting variable acceleration moving target.Finally,the effects of different radar working conditions,target motion characteristics and sea clutter characteristics on the sparse representation of sea clutter suppression are analyzed.The operating conditions of the algorithm are given to provide reference for practical engineering applications.The CSIR measured data set verification shows that the output SCR of the proposed method is improved by 3?7dB compared with the traditional suppression method,and the target detection performance is significantly improved.3)The marine small target detection method based on deep convolutional network is studied.The radar echo signal is constructed into target background images suitable for deep learning network size and convolution operator morphology.A training set and test set containing 8455 samples are established,and the input sample set is expanded by translation,flipping,etc,which can verify algorithm performance and improve network generalization capabilities.According to the difference of deep network characteristics between marine target echo and sea clutter,the Faster R-CNN deep convolutional network architecture is improved.And the batch size is added to realize the speed and accuracy improvement,and the temporary inactivation(Dropout)is utilized to prevent over-fitting,which enhances the generalization performance and robustness of the network.The improved deep convolution network can automatically detect the marine target,realizing the transition from the marine target detection method based on statistical model to the data-driven detection method.The experiments on CSIR dataset show that when the false alarm rate is 10-3,the detection probability is 58%,which is 28%higher than traditional CFAR detection.The new direction of deep learning sea surface target detection is explored.