Study On High-Dimensional Feature Space Learning-Based Detection Methods Of Sea-Surface Small Targets

It is a difficult task to detect sea-surface small targets for sea-surface surveillance radars.The reasons are as follows.First,sea clutter collected by high-resolution radars has strong non-Gaussian characteristics,which is difficult to describe with accurate statistical models,and high-resolution mode also aggravates the influence of sea spikes.Second,small targets have small RCSs and their returns have low SCRs,which makes it difficult to detect them in the background of strong clutter.Third,considering the diversity of targets,it is impossible to collect all kinds of interested target returns.Fourth,the interactions between targets and sea waves lead to complex amplitude and Doppler modulations of the target returns so that it is difficult to use statistical models to describe target returns in all sea states.So high spatial resolution and long observation time are two commonly-used technical means.The former lowers sea clutter level and the latter increases integration of target returns.Long observation time is often implemented by fast scan mode in anti-submarine radars and beam dwelling mode in ubiquitous radars,where the long observation time is achieved by radar multi-scanning cycles and beam dwelling in the former and latter respectively.Because sea clutter time series and target returns fail to be modelled by simple parametric models,it makes sense to develop detectors using features based on intuition.However,traditional feature-based detectors not only face the high computational cost,but also suffer from the dimension limitation.These conventional detectors cannot allow the use of more than three features,which will seriously affect the performance improvement of the detectors.In order to solve abovementioned problems,some detectors which can directly work in high-dimensional feature space for improving detection performance are developed.Meanwhile,some dedetectors which have fast decision for real-time processing of radars are developed.The core research contents can be summarized as follows.1.In terms of the lack of phase information in traditional fractal-based detectors,a detection method based on the all-dimensional Hurst exponent is proposed for detecting floating small targets on the sea.In this method,nonstationary sea clutter time sequence,and corresponding amplitude sequence and phase sequence are modeled as statistically similar fractional-order Brownian motion processes at different time scales.Their Hurst exponents form an all-dimensional description of fractal characteristics of sea clutter time sequence.Radar returns with targets and sea clutter exhibit salient differences in the all-dimensional description.Averaging three Hurst exponents followed by an adaptive threshold decision gives a simple but effective detector of sea-surface small targets.Experimental results on open and recognized radar databases show that the proposed detector has low computational cost and achieves better performance than detectors using only amplitude features.2.In terms of the dimension limitation of existing feature-based detectors,a small target detection method based on feature compression is proposed.In this method,a generator of simulated target returns is firstly given to balance the numbers between sea clutter and target returns samples for subsequent detector design.Secondly,a feature compression algorithm based on the maximization of inter-class distance is proposed,which can compress high-dimensional samples into three-dimensional ones.Finally,a improved convexhull learning algorithm is presented,and the directional guidance of its decision region is carried out by using the simulated target returns.Using the compressed three-dimensional samples as the inputs of the improved convexhull learning algorithm,a detector with controllable false alarm rate can be obtained.Verified by open and recognized radar datasets,the feature compression-based detector effectively improves the detection probability and robustness.3.In terms of the unavoidable compression loss and the limitation of performance improvement caused by the convex hull in the feature compression-based detector,a sea-surface small target detection method based on anomaly detection and K-Nearest Neighbor(KNN)is proposed.In this method,a non-convex preliminary decision region is first constructed by anomaly detection theory,which is determined by the hyper-spherical coverage of the sufficient and ergodic sea clutter training set.If the test sample does not fall into the preliminary decision region,it is seen as target directly;otherwise,it enters the KNN-based classifier.The KNN-based classifier obtains the test statistics through feature weighting,neighborhood weighting and distance weighting,and then completes the detection by threshold decision.The detection method can be directly designed in the high-dimensional feature space,which effectively avoids the compression loss,and also reduces the performance loss caused by the mismatch between the decision region and samples distribution.Finally,verified by open and recognized radar datasets,the KNN-based detector has significant performance improvements.4.In terms of the complex distance calculation between high-dimensional samples in the KNN-based detector,a sea-surface small target detection method based on a decision tree with pre-Gini index is proposed.In this method,to preliminarily realize the rough control of the false alarm rate,a preferential decision tree(pre-decision tree)is proposed,where a preferential Gini index(pre-Gini index)is defined to replace the Gini index and control the false alarm rate roughly.Then,an improved pruning is added to the pre-decision tree to generate a “dominant clutter tree” based on the anomaly detection framework,which can accurately control the false alarm rate at a desired level.The proposed method can directly work in the high-dimensional feature space,and its decision is simple,which effectively avoids the distance calculation between high-dimensional samples.Through the verification of the open and recognized radar datasets,the proposed detector improves the performance,and the computational cost has been initially reduced.5.In terms of the high computational cost of feature-based detectors in high-dimensional feature spaces,a fast small target detection method based on feature normalization and fusion is proposed.In this method,it is found that the seven features of sea clutter can be modeled well by the Burr type XII and t-distributions.From the fitted distribution,each feature is normalized to approximate to the standard Gaussian distribution by a nonlinear transformation.In the seven dimensional normalized feature space,an analytical method is given to calculate the optimal weights of feature fusion by maximization of the interclass distance of the two-class samples.Owing to the normalization,the normalized feature fusion loss is significantly reduced comparing with the direct fusion of seven features.Verified by open and recognized radar datasets and a UAV dataset,the fast small target detection method based on feature normalization and fusion not only greatly reduces the computational cost,but also achieves good generalizability and comparable detection performance.