| 摘要: |
| 合成孔径雷达（Synthetic Aperture Radar，SAR）图像中舰船小目标尺寸小、分辨率低、特征提取难，容易出现漏检和误检导致识别准确率不高。为此，提出改进YOLO-v7网络以增强对小目标特征的提取，从而提高识别检测准确率。针对采用YOLO-v7进行SAR图像船舰小目标检测过程中特征通道信息缺失的问题，通过在主干特征提取网络中引入了坐标注意力机制(Coordinate Attention，CA)，避免了输入特征在提取过程中位置的丢失。然后，引入“重脖子”的GraffeDet-Lite网络架构替换颈部网络（Path Aggregation Network，PANet）结构来获得更细粒度和更强化的特征，提高多尺度特征利用率和检测精度；接着，在检测头中采用高效交并损失比函数（Efficient Intersection over Union Loss，EIOU Loss）提升目标定位效果，进而提高SAR舰船图像算法检测精度。通过RSDD-SAR：SAR舰船斜框检测数据集对比实验表明，提出的改进YOLO-v7算法能有效提高SAR图像船舰小目标识别检测能力，漏检情况得到明显改善，识别准确率达到了90.8%，与未改进前的YOLO-v7相比提高了3.2%。 |
| 关键词: SAR图像 舰船目标检测 YOLO-v7 CA注意力机制 GraffeDet-Lite |
| DOI：10.20079/j.issn.1001-893x.240515004 |
|
| 基金项目: |
|
| A Ship Target Detection Algorithm Based on Improved YOLO-v7 |
| MAO Yunlong,GE Rui,WEI Hangcheng,LI Si |
| (1.College of Oceanography,Jiangsu University of Science and Technology,Zhenjiang 12003,China;2.School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 528463,China) |
| Abstract: |
| Due to the small size,low resolution,and difficult feature extraction of small targets in synthetic aperture radar(SAR) images,they are prone to missed and false detections,resulting in low recognition accuracy.Therefore,it is proposed to improve the YOLO-v7 network to enhance the extraction of small target features,thereby improving the accuracy of recognition and detection.In response to the problem of missing feature channel information in SAR image ship small target detection using YOLO-v7,coordinate attention mechanism is introduced in the backbone feature extraction network to avoid the loss of input feature position during the extraction process.Then,the “heavy neck” GraffeDet Lite network architecture is introduced to replace the neck network PANet structure to obtain finer grained and stronger features,improving multi-scale feature utilization and detection accuracy.Next,the EIOU Loss function is used in the detection head to improve the target localization effect,thereby improving the detection accuracy of the SAR ship image algorithm.The comparative experiment on the RSDD-SAR:SAR ship oblique frame detection dataset shows that the proposed improved YOLO-v7 algorithm can effectively improve the recognition and detection ability of small targets on ships in SAR images.The missed detection situation is significantly improved,and the recognition accuracy reaches 90.8%,which is 3.2% higher than that of the unmodified YOLO-v7. |
| Key words: SAR images ship target detection YOLO-v7 coordinate attention mechanism GraffeDet Lite |
《电讯技术》编辑部 