ABSTRACT

Remote sensing data generated by satellites have important research and application values. However, traditional detection models have poor adaptability and generalization ability. In order to solve the low accuracy problem of the existing algorithm for small object detection of remote sensing, a small object detection algorithm MARNet (multi-angle rotation network) for remote sensing images of multi-angle rotation was proposed in this study, which used ResNet101 (residual network) as the baseline network. Global attention feature pyramid networks (GA_FPN) structure was designed based on the features of the pyramid network to improve the small object detection performance in remote sensing. Then MergeNet (Merge Network) was designed to better obtain the semantic relationship between features, and the attention mechanism was introduced to enhance the feature information of the target object. Datasets of DOTA (a large-scale dataset for object detection in aerial images) and NWPU VHR-10 (northwestern polytechnical university, very-high-resolution) are used to verify the algorithm.

摘要

卫星产生的遥感数据具有重要的研究和应用价值。但是，传统的检测模型具有较差的适应性和泛化能力。为了解决现有的遥感小目标检测算法精度低的问题，提出了一种用于多角度旋转遥感图像的小目标检测算法MARNet（多角度旋转网络），该算法用于ResNet101（残留网络）作为基准网络。基于金字塔网络的特征设计了全球注意力特征金字塔网络（GA_FPN）结构，以提高遥感中的小物体检测性能。然后设计MergeNet（合并网络）以更好地获取要素之间的语义关系，引入了注意力机制来增强目标对象的特征信息。使用DOTA（用于在空中图像中进行对象检测的大规模数据集）和NWPU VHR-10（西北工业大学，非常高分辨率）的数据集来验证该算法。