面向自适应目标检测的训练样本筛选方法研究现状与进展

Screening Approaches of Training Samples for Adaptive Detection:A Review

摘要:

在雷达目标自适应检测中,常假定训练样本与被检测单元中杂波分量独立同分布。实际应用中,各种非均匀因素易导致部分训练样本受异常值污染,因此,研究合适的训练样本筛选方法具有重要意义。将训练样本筛选方法分为 4类,主要包括杂波功率法、知识辅助法、特征相似度法和其他筛选方法。其中:基于杂波功率的筛选方法可细分为功率选择训练、广义内积和自适应功率剩余等；基于知识辅助的筛选方法主要分为两类,一类是利用信息直接筛选地形数据,另一类则对检验协方差矩阵进行重构；特征相似度方法主要研究训练样本间频谱和协方差矩阵的差异性；其他筛选方法侧重于融合稀疏恢复、子孔径、多帧处理和深度学习等技术进行训练样本筛选。最后,对训练样本筛选方法进行总结和展望,给出现有研究在实际应用中所面临的潜在问题和挑战。

Abstract:

In adaptive radar target detection, a common assumption is that the training samples and the clutter components within the cell under test follow an independent and identical distribution. However, in real-world applications, a multi-tude of nonhomogeneous factors can cause the contamination of some training samples by outliers. Consequently, devel-oping effective training sample selection techniques is crucial. First, these techniques are categorized into four main types based on existing literature: clutter power methods, knowledge-aided methods, feature similarity methods, and other selec-tion methods. Clutter power methods can be divided into power selection training, generalized inner product, and adaptive power residual. Knowledge-aided methods consist of two types: the first directly selects terrain data using information, while the second concentrates on reconstructing the test covariance matrix. Feature similarity methods focus on the dis-crepancies between the spectral and covariance matrices of training samples. Other methods emphasize the integration of various technologies with training sample selection, including sparse recovery, subaperture processing, and multi-frame techniques. In conclusion, the training sample selection methods are summarized and prospected, the potential challenges and issues faced by the screening methods in practical applications are given.