基于多模型协同的舰船目标检测

Multiple models collaboration for ship detection

随着遥感成像技术的不断进步,遥感图像的舰船目标检测已成为确保海上运输安全和效率的关键手段,对海上交通、环境保护及国家安全至关重要。然而,由于舰船目标尺度差异大、背景复杂等问题,现有单一检测模型的方法过度依赖训练数据,无法适应尺度多变的舰船目标。提出了一种多模型协同训练的框架,利用多个已训练好的舰船检测模型作为辅助网络,通过知识迁移的方式辅助优化目标数据的主网络。首先,通过三元关系约束建立辅助网络与主网络间的分布知识传递;其次,采用软标签引导策略整合辅助网络中的标签知识,提高舰船检测的准确性。实验结果表明:相较于现有主流方法,所提方法在DOTA和xView数据集上展示了较好的性能,克服了单一模型的局限性,为遥感图像的目标检测提供了新的解决思路。

In the context of globalization,the importance of ship monitoring is becoming more and more prominent.With the continuous progress of the remote sensing imaging technology,ship detection has become a key means to ensure the safety and efficiency of maritime transportation,and is crucial for maritime transportation,environmental protection,and national security.However,due to the large difference in scales and complex background of ship targets,existing single detection model methods rely too heavily on training data and cannot adapt to ship targets with variable scales.In this paper,we propose a multiple models collaboration framework,in which multiple trained ship detection models are regarded as auxiliary network,and the main network training is optimized by knowledge migration.First,ternary relationship constraints are introduced to transfer knowledge between the auxiliary network and the main network.Then,a soft-label guidance strategy is proposed to further improve the accuracy of ship detection.The experimental results show that compared with the existing mainstream methods,the proposed method demonstrates better performance on DOTA and xView datasets,overcoming the limitation of a single model and providing a new solution idea for target detection in remote sensing images.