太赫兹科学与电子信息学报2025，Vol.23Issue(11)：1141-1149,1156,10.DOI:10.11805/TKYDA2024401
基于改进YOLOv8网络模型的SAR图像船舰小目标检测算法
An algorithm for detecting small ship targets in SAR images based on the improved YOLOv8 network model
张珺珲 1李勃 1赵泽楷1
作者信息
- 1. 南京航空航天大学 电子信息工程学院,江苏 南京 210016
- 折叠
摘要
Abstract
To address the problems of numerous small targets,insufficient feature acquisition,and the low proportion of detectable information in Synthetic-Aperture-Radar(SAR)images,an innovative refinement of the standard YOLOv8 architecture is presented.Firstly,all original convolutional blocks are replaced by Spatial-to-Depth Convolution(SPD-Conv)layers that preserve fine-grained spatial detail.Secondly,a Bidirectional Feature Pyramid Network(BiFPN)augmented with an extra small-target detection layer is introduced;its bi-directional feature flow markedly improves both the efficiency and the effectiveness of multi-scale fusion,strengthening the model's grasp of local details and global context.Thirdly,a Convolutional Block Attention Module(CBAM)is embedded to boost representational power while keeping the computational overhead low.Extensive experiments on the HRSID and SSDD benchmarks show that the enhanced YOLOv8 raises mAP@0.5 by 1.2%and 0.5%,respectively,trims the parameter count by 32.6%,and increases Giga Floating-point Operations Per Second(GFLOPs)by 104.9%,verifying its practical value for small-target detection in SAR imagery.关键词
合成孔径雷达(SAR)/图像检测/空间深度转换卷积(SPD-Conv)/小目标检测层/双向特征金字塔网络(BiFPN)/卷积注意力机制(CBAM)Key words
Synthetic Aperture Radar(SAR)/image detection/Spatial-to-Depth Convolution(SPD-Conv)/small-target detection layer/Bidirectional Feature Pyramid Network(BiFPN)/Convolutional Block Attention Module(CBAM)分类
信息技术与安全科学引用本文复制引用
张珺珲,李勃,赵泽楷..基于改进YOLOv8网络模型的SAR图像船舰小目标检测算法[J].太赫兹科学与电子信息学报,2025,23(11):1141-1149,1156,10.
