| 摘要: |
| 海上目标检测中，传统单一特征检测方法在面对复杂场景时往往表现不佳。为了进一步提升检测性能，本文提出了一种基于改进无监督自适应嵌入（Unsupervised Adaptive Embedding, UAE）的海上目标检测方法。该方法从时域、频域、时频域等多个变换域挖掘了海杂波和目标回波的差异特征，构建了一个六维初始特征空间。在此基础上，通过改进UAE方法，在降维过程中引入正则化马氏（Mahalanobis）距离来精确衡量特征点之间的相似性，并利用类间和类内散布矩阵优化数据的判别性，通过特征投影矩阵将六维初始特征空间映射到三维特征空间，保留关键信息的同时实现数据的压缩。应用凸包学习实现了三维判决区域的构建与目标检测。在海军航空大学雷达对海探测数据集进行了性能验证，结果表明，相较于已有三特征检测方法，所提方法在检测概率方面具有显著的性能提升。 |
| 关键词: 特征提取 小目标检测 海杂波 凸包学习 |
| DOI：DOI:10.3969/j.issn.1672-2337.2025.04.009 |
| 分类号:TN957.51 |
| 基金项目:国家自然科学基金资助项目（No.62388102， 62101583）； 烟台市2023年校地融合发展项目（No.2323013?2023XDRH001）； 山东省科技型中小企业创新能力提升工程计划项目（No.2023TSGC0823）； 烟台市科技型中小企业创新能力提升工程计划项目（No.2023TSGC112） |
|
| Detection Algorithm of Small Target on the Sea Surface Based on Unsupervised Adaptive Embedding |
|
ZHAO Guomiao， DING Hao， CAO Zheng， HE Pengfei
|
|
1. School of Physics and Electronic Information, Yantai University, Yantai 264005, China;2. Naval Aviation University, Yantai 264001, China
|
| Abstract: |
| In maritime target detection, traditional single?feature detection methods often struggle in complex scenarios. To enhance detection performance, a maritime target detection approach based on improved unsupervised adaptive embedding （UAE） is introduced in this paper. This approach distinguishes between sea clutter and target echoes by extracting distinctive features from multiple transform domains, including time domain, frequency domain, and time?frequency domain, forming an initial six?dimensional feature space. Building on this, the UAE method incorporates the regularized Mahalanobis distance for precise similarity measurement during dimensionality reduction, and leverages inter?class and intra?class scatter matrices to optimize data separability. The feature projection matrix is then employed to map the six?dimensional feature space into a three?dimensional space, compressing data while preserving critical information. Using convex hull learning, a three?dimensional decision region is constructed for target detection. Performance validation on the radar sea detection dataset of Naval Aviation University demonstrates that the proposed method significantly improves the detection probability compared to existing three?feature detection approaches. |
| Key words: feature extraction small target detection sea clutter convex hull learning |
