Improved Methods For Detecting Small Targets On Sea Surface Based On Classification Strategy

An important task of seaward radar is to detect targets on sea surface.Under the influence of complex environment,the detection of small targets at sea clutter such as buoys,ice floes,small vessels,etc.has been an important challenge for traditional target detection algorithms.Feature-based detection methods have been widely used in the detection of small targets at sea,the mainstream algorithms are based on single-classification convex hull learning algorithm and binary classification-based pattern recognition algorithm,convex hull learning algorithm is limited to the feature space of no more than 3 dimensions,once the number of dimensions increases,the computational volume spikes can not be quickly converged.The pattern recognition algorithm based on binary classification faces the dilemma of non-equilibrium in the category.Traditionally,the problem is solved by adding artificial simulation of target echoes.But this step will increase the complexity of the algorithm and the detection results are easily affected by the simulation accuracy.In this paper,based on the framework of anomaly detection,we propose an improved multi-feature detection algorithm with self-encoder(AE)and isolated forest(i Forest)to break the feature dimension limitation.Based on the idea of multi-classification,the class non-equilibrium problem of the binary classification algorithm is solved by using multi-classifiers jointly.The main research contents are as follows:1.A multi-feature detection method based on improved AE is proposed,which breaks through the feature dimension limitation.The algorithm is based on the anomaly detection framework,and the AE neural network is trained by inputting a purely clutter features set in the training phase,which can get the reconstruction error set,and the detection threshold is obtained by Monte Carlo method.The category verdict is derived by calculating the reconstruction error of the test sample and the detection threshold in the testing phase.Experimentally verified,the algorithm improves the detection performance compared with the traditional single classification convex hull learning algorithm.2.A multi-feature detection algorithm based on the improved i Forest is proposed.Also based on the anomaly detection framework,the i Forest is trained by inputting a purely clutter features set in the training phase,and the set of anomaly indices can be obtained,and the detection threshold is determined using Monte Carlo method.The category verdict is derived by calculating the anomaly indices of the test samples in the testing phase compared with the detection threshold.Experimentally verified,this algorithm has better detection performance and stability compared with the convex hull learning algorithm and the improved AE-based algorithm.3.A multi-feature detection algorithm based on the idea of multi-classification is proposed to solve the non-equilibrium problem of sea clutter and target echo category.The feature space is constructed by first extracting multiple features from the sea clutter data and the target data,and then dividing the sea clutter feature space into multiple subspaces based on the multi-classification "one-to-one" method,and constructing multiple binary classifiers for joint judgment.A two-parameter K-NN algorithm is designed for each sub-classifier to achieve adjustable false alarm rate.After experimental verification,this algorithm is compared with the K-NN-based detection algorithm,which does not require manual simulation of target echoes and has a certain performance improvement.