| 摘要: |
| 本文针对海面雷达弱目标在强非高斯海杂波中难以检测的问题,提出一种融合变分模态分解(VMD)、傅里叶同步压缩变换(FSST)与改进ResNet50的检测方法。针对IPIX海杂波数据集信号,先通过VMD分解自适应判别目标与杂波距离门,筛选最优特征模态重构信号以抑制杂波并规避模态混叠；再经FSST时频重分配生成时频图,强化目标时频特征；最后构建融入深度可分离卷积、SimAM注意力及h-swish激活的改进ResNet50网络,输入时频图完成目标检测。大量实验表明,本文方法的目标检测准确率显著优于对比方法。且改进后模型参数量、浮点运算量(FLOPs)较原始ResNet50分别缩减47.97%、96.42%,推理速度提升46.21%,实现检测性能与轻量化、实时性的最优平衡,为海杂波背景下弱目标高效检测提供可靠技术支撑。 |
| 关键词: 海杂波 弱目标检测 时频分析 ResNet 注意力机制 |
| DOI： |
| 分类号:TN957.51 |
| 基金项目:国防基础科研稳定支持专题项目（JCKYS2020604SSJS010） |
|
| Sea-Surface Weak Target Detection Based on VMD-FSST Time-Frequency Analysis and Improved ResNet50 |
|
尚尚
|
|
|
| Abstract: |
| To address the difficulty of detecting weak radar targets in strong non-Gaussian sea clutter, this paper proposes a method combining Variational Mode Decomposition(VMD), Fourier Synchrosqueezing Transform (FSST),and an improved ResNet50.First, VMD adaptively identifies target and clutter range gates in IPIX radar signals, reconstructing the signal with selected modes to suppress clutter. FSST then enhances the time-frequency representation. Finally, a lightweight ResNet50 incorporating depthwise separable convolution, SimAM attention, and h-swish activation performs detection on the time-frequency maps. Experiments show superior accuracy over existing methods,while reducing parameters and FLOPs by 47.97% and 96.42%, and improving inference speed by 46.21%,achieving an effective balance between performance and efficiency. |
| Key words: sea clutter weak target detection time-frequency analysis attention mechanism |