Research On Detection And Classification Of Sea-surface Small Targets Based On Machine Learning

The detection and classification of floating small targets in sea clutter are two difficult points in the field of sea surface search radar surveillance,which have important guiding significance for marine land protection,maritime assistance,maritime affairs and channel management.In the increasingly refined radar system and target miniaturization environment,the performance of traditional detection and classification methods based on sea clutter statistical modeling and critical signal-to-clutter ratio are very limited.Artificial intelligence-led machine learning technology and its important branch deep learning technology have excellent nonlinear fitting and feature extraction capabilities,which provide new ideas for breaking through performance bottlenecks.Therefore,with the help of machine learning technology,this thesis aims to develop new target feature detection and classification methods that can automatically perceive sea clutter characteristics and match the multi-dimensional characteristics of the target and sea clutter as well as differential recognition.The main research content of the thesis is divided into three parts:(1)A classification algorithm for floating small targets based on migration learning is proposed.Aiming at the problems of unbalanced sea clutter and target categories,and the scarcity of target samples,while taking into account the advantages of micro-Doppler in analyzing the micro-motion characteristics of floating small targets and the advantages of convolutional neural networks in image processing,general low-dimensional hidden features are extracted from data sets of different scenes based on transfer learning and reused on training of convolutional neural networks for target classification.The classification results on the measured data show that the network training based on migration learning converges faster,and the classification performance is better than the network training without lowdimensional features.(2)A target detection algorithm based on the feature of connected regions is proposed.The visual difference between sea clutter and the target on the time-frequency diagram is observed,which is specifically manifested in the difference in the shape distribution of the energy clusters.Based on the clutter block whitening process,the sea clutter energy is suppressed.Then feature of connected regions is extracted by the pixel-level classification network to realize target detection.The detection results based on the measured data show that the block whitening process can suppress the clutter to a certain extent,and the performance of the designed detector is better than the traditional fractal-based detector.Finally,selection of detection statistics and the training method of network are optimized,which improve the detection performance further.(3)A multi-feature joint floating small target detection algorithm is proposed.Performance of traditional feature detection methods is limited by the feature dimension.For example,the convex hull learning algorithms often used in anomaly detection have very limited processing of high-dimensional feature vectors,which is generally difficult to break through the three-dimensional.However,as the cognition of sea clutter and the target is deepened and gradually increased,the problem of low feature utilization will be encountered.In addition,the computational complexity of most feature fusion algorithms is relatively high,which is not conducive to engineering implementation.In order to break through the performance bottleneck and improve the feature utilization rate,a multi-feature joint detector based on the integrated learning idea is proposed.The separability of different features is analyzed and the decision tree is used to implement joint decision-making to improve the robustness of the detector.The detection results based on the measured data show that the performance of the designed detector is improved compared with the traditional detectors.