基于对比学习的航海雷达目标检测方法

Marine Radar Object Detection Method Based on Contrastive Learning

由于航海雷达图像中的目标与杂波的相似度较高,因此目标检测任务非常困难.此外,虽然航海雷达的原始数据量很大,但标注需要大量的专业知识,导致目前可以直接使用的有效数据很少.为解决上述问题,本文首先建立了两个航海雷达数据集,分别是无标签的航海雷达数据集(Unlabeled Marine Radar Dataset,UMRD)和有标签的航海雷达检测数据集(Marine Radar Detection Dataset,MRDD).同时,本文提出了一种基于对比学习的航海雷达目标检测方法(Contrastive Learning for Marine Radar Detection,CLMRD).该方法首先以聚类的方式产生伪标签,然后以交替预测的方式从样例级别提高特征的判别性,并根据一致性准则从数据分布级别提升特征判别性.接下来,使用Yolov5作为目标检测网络,并结合预训练的特征提取器进行微调.最后,CLMRD对不同切片的检测结果进行融合.提出的方法在MRDD数据集上达到了0.97的准确率和0.95的召回率,显著优于其他检测方法,验证了其有效性和鲁棒性.

The task of detecting targets in marine radar images is challenging due to high similarity between the tar⁃get and clutter.Although there is a large amount of raw data available for marine radar,annotating them requires expert knowledge,making labeled data particularly valuable.To address these issues,this paper establishes two marine radar data⁃sets,the unlabeled marine radar dataset(UMRD)and the labeled marine radar detection dataset(MRDD).To improve the feature discriminability of the data,this paper proposes a contrastive learning approach for marine radar detection(CLMRD),which involves generating pseudo labels by clustering and then improving the feature discriminability at both the sample and data distribution levels using a consistency criterion.The object detection network Yolov5 is used to detect targets,and fine-tuned with a pre-trained feature extractor.CLMRD fuses the detection results of different slices to improve the accuracy and recall rates.The proposed method achieves an accuracy rate of 0.97 and a recall rate of 0.95 on the MRDD dataset,outperforming other detection methods and demonstrating its effectiveness and robustness.