Intelligent Detection Method For Weak Targets On The Sea Surface Radar Based On Deep Learning

Marine resources are abundant and precious to our country.In the complex and nonuniform sea-clutter environment,detecting weak targets on the sea surface can provide accurate and comprehensive monitoring and protection,improve the detection accuracy and range of marine radar,and have important strategic significance for our country.However,the marine environment is complex and diverse,with high clutter background intensity and target diversity.Existing detection algorithms are difficult to adjust specifically,and there are issues such as decreased detection performance and high false alarms in low signal-to-clutter ratio(SCR)conditions.Therefore,the key issues of radar detection include the detection of weak targets in strong sea clutter backgrounds,and it is challenging and has practical application significance.Based on the above difficulties in non-uniform sea-clutter detection,this thesis uses algorithm based on deep learning which is adopted to build a network and learn parameters,and the detection of weak targets by radar under the background of sea clutter is studied.This thesis conduct in-depth research on clutter suppression and target detection algorithms,and use simulated targets based on actual measurement data to verify performance and improve radar target detection.The main work and research contents of this thesis are as follows:1.This thesis studied the statistical distribution model of sea clutter,and uses model fitting based on actual measurement data,this thesis explained that there is currently no accurate mathematical model that can describe all clutter distributions.Based on actual measurement data,this thesis used singular value decomposition to extract different feature subspaces,and studied the clutter suppression method based on fractional Fourier transform(SVD-FRFT)in the fractional order domain,which can highlight the target signal to a certain extent on actual measurement data.2.In view of the limited clutter suppression effect of traditional methods and the difficulty in applying them in radar scanning mode,this thesis studied the objective function based on Wasserstein distance and proposed the clutter suppression network based on Wasserstein GAN Gradient Penalty(WGAN-GP),which can ensure the stability of network training during adversarial game parameter training.Combining U-Net pixellevel clutter suppression and Patch-GAN discriminator,this thesis constructed a WGANGP-based training and testing network architecture,designed network parameter models with different SCR ratios on actual measurement data,analyzed and compared the suppression ability of different models,and verified the effectiveness and generalization ability of the proposed method.3.This thesis studied the target detection method based on aggregate residual network,which alleviating the effects of the deep network layer due to gradient,leading to decreased parameter training performance.Based on the classic residual network design,this thesis introduced the "cardinality" dimension and increased the parallel number of channels to improve the training and testing effect of the network without changing the network layer.At the same time,using group convolution to design the network model ensures that the training parameters do not explode,thereby reducing the network computation.After the actual measurement data was processed by the WGANGP clutter suppression network,it was input into the aggregate residual network for target detection.After verification on the test set,the proposed method showed a significant improvement in detection probability under low SCR compared with the residual network.The proposed methods in this thesis use actual measurement data as the clutter background,and the experimental results verify that the proposed algorithm can effectively improve the radar’s detection performance of low SCR targets under nonuniform sea-clutter.