Graph-feature Based Detection And Classification Of Small Targets On Sea Surface

Radar surface target detection is a great significance both in military and civilian field.Due to the complexity of the marine environment and the uncertainty of sea clutter,it brings large challenges to the detection of floating small targets on the sea surface.Traditional statistical target detection methods that perform function modeling based on the characteristics of sea clutter will encounter problems such as high false alarms and impaired detection performance.With the development of signal processing technology,the method of extracting different features is used to detect targets.Application of graph structure in the fields of biological genes,image processing,big data,etc.Uses the characteristics of graph structure and and the non-linear,non-Gaussian,and non-stationary characteristics of sea clutter,extract the small target detection algorithm on the sea surface based on graph domain features.In recent years,with the rapid development of deep learning technology,many scholars have introduced deep learning into the classification of sea clutter and targets.In this thesis,we explored different features of graph structure characteristics,and discuss the effectiveness of the research on the detection and classification of small sea targets based on the graph features.Based on the analysis of sea clutter characteristics and graph structure,we further study that correlation and connectivity density of radar echo data,the maximum eigenvalue of graph Laplacian matrix and graph entropy are proposed as features to detect sea surface target respectively.The detectors are analyzed and completed.Furthermore,the classification algorithm of sea clutter and target is proposed by combining the graph structure with the convolutional neural network.The research content of this thesis is summarized as follows:(1)The definition of graph structure and the relationship and characteristics of data are introduced.After preprocessing the measured data,the graph topology is established and the adjacency matrix is generated according to the graph connectivity density obtained from the correlation of the data.The difference between the topology structure of the clutter data graph and the topology structure of the target data graph is analyzed.And we lay the foundation for the feature extraction and classification of the graph below.(2)Explore the correlation and connectivity density of radar echo data in the frequency domain to reduce the loss of echo energy.Firstly,according to the graph theory,the degree matrix is constructed by the adjacency matrix,and the Laplace matrix is obtained.Then,according to the relationship between the eigenvalues of the graph Laplacian matrix,the maximum eigenvalue of the graph Laplacian matrix is extracted as the target detection feature,and the sea surface target detection is completed.Finally,the experimental results of the maximum eigenvalue of the graph Laplacian matrix have superior detection performance and less computational complexity than some existing feature algorithms.(3)According to the data distribution of the adjacency matrix,the aggregation and isolated points of the clutter data and the target data,the target detection is realized.After the isolated points are removed from the adjacency matrix,a new vector is formed,and its entropy is calculated,the sea surface target detection based on graph entropy is proposed.At the same time,the graph entropy is compared with the other entropies in different domains,and the detection performance is significantly improved.(4)Further we combined graph structure with Convolutional Neural Networks(CNN)in deep learning,and proposed a sea surface target map classification based on Convolutional Neural Networks.The graph constructed by the adjacency matrix is trained on the model.Through the analysis of various indicators,the effectiveness of using CNN to classify the sea surface target graph is verified,and we compared with other machine learning methods.The experimental results show that this method can better classify the sea clutter and target.