Enhanced local sparsity coefficient-based sea-surface floating target detection

Floating Small Target Detection in Sea Clutter Based on Multifeature Angle Variance

“…12 with the false alarm rates change from 0.001 to 0.1. In the several detectors, the LDOF-based detector > REPLACE THIS LINE WITH YOUR MANUSCRIPT ID NUMBER (DOUBLE-CLICK HERE TO EDIT) < [7] (c) Detection results of modified phase-feature-based detector [9] (d) Detection results of TF-feature-based detector [8] (e) Detection results of LOF-based detector [15] (f) Detection results of isolation-tree-based detector [15] (g) Detection results of KNN-based detector [14] (h) Detection results of proposed ABOD-based detector [21] relies on the distance of the feature samples, and the LOF-based detector [15] and the ELSC-based detector [22] rely on the feature sample density in the high-dimensional feature space. Under four polarizations, it is evident that the detection performance of the proposed ABOD-based detector is always optimal, and the performance improvement is more pronounced at lower false alarm rates, making it more suitable for practical applications for radar target detection at low false alarm rates.…”

“…In the final experiment, advantages of using the angles as the measure of the anomaly detectors are validated. The detection performance of the density-based detectors (LOF-based detector [15] and ELSC-based detector [22]), the distance-based detector (LDOF-based detector [21]) and the angle-based detector (proposed ABOD-based detector) are compared. For the sake of fairness, the several detectors use the same eleven features presented in the section II.…”

“…Therefore, in order to combine multiple features and break the dimension limitation, amounts of machine learning algorithms have been modified and applied to the feature-based detection [10-12, 14-15, 20-23]. Thereinto, some algorithms only use the normal samples to train the decision region, such as the one-class support vector machine (OCSVM) learning algorithm [20], the isolation forest (iForest) learning algorithm [15], the local outlier factor (LOF) learning algorithm [15], the local distance-based outlier factor (LDOF) learning algorithm [21], the enhanced local sparsity coefficient (ELSC) learning algorithm [22]. And some algorithms, such as the K-nearest neighbour (KNN) learning algorithm [14], and the fast principal component analysis (PCA) learning algorithm [23].…”

“…The approxMABOF obtained in (22) is distinguishable between returns with target and sea clutter, and is suitable to be the test statistic for the ABOD-based detector, and then the construction of the detector can be divided into two branches as shown in Fig. 7.…”

“…In this part, the eleven features mentioned above are extracted from a large amount of pure sea clutter time sequences then a sea clutter feature vector set is obtained which is used as training sample set S0 under the null hypothesis H0. Then the approxMABOF of the feature vectors in S0 are calculated by (22), and an improved fastABOD model is constructed. Because of the need to accurately control the probability of false alarms, the Monte Carlo method is often referred to as the detection threshold when the false alarm rate is given.…”