海杂波中基于可控虚警K近邻的海面小目标检测

由于高分辨海杂波具有复杂的特性以及海面小目标具有多样性，没有精确的简单统计模型可以较好地描述海杂波和目标回波时间序列，这导致目标检测遇到了很多阻碍。为了区分海杂波和目标回波，分别提取它们的特征将检测问题转化为特征空间中的分类问题是一种有效的方法。基于特征的检测可以归结为在特征空间中的一种2元假设检验问题，但是其有两个问题需要解决:一是目标回波数据远少于杂波数据；二是虚警概率不可控。为了解决第1个问题，一种典型小目标的仿真回波产生器被用于产生充足的典型目标回波数据，以辅佐后续检测器的设计。K近邻(K-NN)是一种简单有效的分类方法，但是因为无法精确地控制虚警率而不能直接在目标检测中使用。该文提出一种基于改进K-NN的海面小目标检测方法，可以很好地实现可控虚警。经IPIX雷达数据集验证，所提出的方法在观测时间分别为0.512 s和1.024 s时获得了85.1%和89.2%的检测概率，相比现有的检测器获得了7%和5%的提升，具有良好的检测效果和稳定性。 

Sea-surface Small Target Detection Based on K-NN with Controlled False Alarm Rate in Sea Clutter

Owing to the complicated characteristics of high-resolution sea clutter and the diversity of sea-surface small targets, there is no precise parameter model to describe sea clutter and returns with targets. As a result, target detection faces many obstacles. To distinguish sea clutter and target returns, it is effective to extract their features to transform the detection problem into a classification problem in feature space. Feature-based detection is a binary hypothesis test in the feature space that encounters two intrinsic difficulties: one difficulty is insufficient target returns versus sufficient sea clutter; the other difficulty is an uncontrolled false alarm rate in detection. To solve the first difficulty, a generator of typical targets returns that can generate sufficient simulated targets returns is used to balance the number of samples between two classes and assist to design the detector. K Nearest Neighbors (K-NN) is the type of classification method that is simple and effective; however, it cannot be used to detect small targets directly because of the uncontrolled false alarm rate. This paper proposes a modified K-NN method with a controlled false alarm rate for detecting small targets. Experimental results on the IPIX radar database indicate that the proposed detector attains 85.1% and 89.2% rates of target detection for the observation time of 0.512 s and 1.024 s, respectively, compared with other existing feature-based detectors, the proposed detector exhibits 7% and 5% improvement, respectively. Thus, the proposed detector exhibits more stable and effective detection performance than other existing feature-based detectors.