| 摘要: |
| 为解决复杂海况下海面小目标检测方法性能不佳的问题，本文提出了一种基于主正交潜在成分分析与卷积神经网络级联的改进检测方法（POLCA-CNN）。首先，针对传统6维雷达回波特征对动态目标特性表征不足的问题，通过引入幅度熵和多普勒非广延熵等熵特征，构建了14维扩展特征空间，有效提升了特征对目标与海杂波的区分能力。其次，针对高维特征的维度灾难问题，本文提出改进的POLCA-Net特征降维模型，通过重构损失、正交损失、质心损失、分类损失和物理约束损失的五重联合优化机制，实现特征维度至3维的有效压缩。该模型在保留信号物理特性的同时，利用对比学习增强目标与杂波的特征可分性。最后，设计了一种轻量化的CNN分类网络，通过可调阈值实现虚警概率的控制。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，所提的改进POLCA-CNN检测器在积累时间为0.512s，虚警概率为0.001时，平均检测概率达到0.861，具有良好的检测效果，为复杂海况条件下的雷达小目标检测提供了有效的解决方案。 |
| 关键词: 海杂波 小目标检测 虚警可控 特征提取 卷积神经网络 |
| DOI： |
| 分类号:TN957.51 |
| 基金项目:国家自然科学基金资助项目、烟台市2023年校地融合发展项目、山东省科技型中小企业创新能力提升工程计划项目、 烟台市科技型中小企业创新能力提升工程计划项目 |
|
| Research on sea surface small target detection method based on POLCA-CNN |
|
|
|
|
| Abstract: |
| To address the issue of poor performance inmaritime small-target detection methods under complex sea conditions, this paper proposes an improved detection method based the cascade of Principal Orthogonal Latent Component Analysis and Convolutional Neural Network (POLCA-CNN). First, to overcome the problem of insufficient representation of dynamic target characteristics traditional 6-dimensional radar echo features, we introduce entropy features such as amplitude entropy and Doppler non-extensive entropy, and construct a 14-dimensional extended feature, effectively enhancing the feature"s ability to distinguish targets from sea clutter. To address the problem of high computational complexity caused by high-dimensional features, this paper proposes an improved POLCA-Net feature dimensionality reduction model. Through a five-fold joint optimization mechanism of reconstruction loss,orthogonal loss,centroid loss,classification loss, and physical constraint loss, effective compression of the dimensions to 3D is achieved. The model retains the physical characteristics of the signal while utilizing contrastive learning to enhance the feature separability of targets and clutter. On this basis a lightweight CNN classification network is designed. The network achieves control of the false alarm probability through a tunable threshold. The performance has been validated using the Naval Aviation University Radar Mar Detection Dataset, and the results show that the proposed improved POLCA-CNN detector achieves an average detection probability of 0.861 at a time accumulation of 0.512s and a false alarm probability of 0.001, demonstrating good detection performance and providing an effective solution for small-target detection of radar under complex maritime. |
| Key words: sea clutter small target detection controlled false alarm rate feature extraction convolutional neural network |