| 摘要: |
| 针对合成孔径雷达（SAR）图像在舰船检测过程中因其环境复杂、舰船尺寸多样而导致检测精度低的问题，提出了一种轻量级的Cross Stage Ghostnetv2(CSG)模块替换YOLOv7网络中的ELAN模块和ELAN-W模块，减少网络的参数数量和计算成本；在颈部网络中引入ConvNeXt模块，加强图片的特征提取能力以提高小目标的检测能力；最后采用K-means++ 聚类自动生成锚框，提高算法的识别精度。在SSDD数据集（SAR Ship Detection Dataset）上的实验结果显示，该算法的平均精度均值mAP@0.5较YOLOv7提升4%，模型参数Params下降10.9%，计算量GFLOPs减少63.7%，在提高精确度的情况下大幅度降低了模型复杂度。实验结果证明了该算法在舰船目标检测上的有效性。 |
| 关键词: 合成孔径雷达图像 舰船目标检测 YOLOv7 |
| DOI：10.20079/j.issn.1001-893x.240319001 |
|
| 基金项目:国家重点研发计划（2023YFC3304900）；中央引导地方科技发展资金项目（YDZX2023050）;山东省海洋环境水下无线传感器网络体系结构研究与设计（2018GGX101038） |
|
| A Ship Target Detection Method in SAR Images Based on Improved YOLOv7 Network |
| HU Huizhong,MAO Yuming,ZHANG Wenkang,DING Qingyan |
| (1.School of Information Science&Electrical Engineering，Shandong Jiaotong University,Jinan 250357,China;2.Shandong Computer Science CenterNational Supercomputer Center in Jinan,Jinan 250014,China) |
| Abstract: |
| For addressing the issue of low detection accuracy in ship detection in synthetic aperture radar(SAR) images due to their complex environment and diverse ship sizes,a lightweight Cross Stage Ghostnetv2(CSG) module is designed to replace the ELAN module and the ELAN-W module in the YOLOv7 network to reduce the number of network parameters and computational costs.The ConvNeXt module is introduced into the neck network to enhance the feature extraction capability of images and improve the detection ability of small targets.The K-means++ clustering is used to automatically generate anchor boxes,and enhance the algorithm’s recognition accuracy.The experimental results on the SAR Ship Detection Dataset(SSDD) show that compared with those of YOLOv7 ,the average accuracy of mAP@0.5 is increased by 4 %,and the model parameters Params is reduced by 10.9 %,and the computational complexity GFLOPs is reduced by 63.7 %,respectively.The algorithm greatly reduces the complexity of the model while improving the accuracy,and is effective in ship target detection. |
| Key words: SAR image ship object detection YOLOv7 |
