基于特征重加权的小样本遥感图像目标检测算法

Few-shot Object Detection on Remote Sensing Images based on Feature Reweighting

摘要:

针对遥感图像具有目标尺度多变、目标模糊、背景复杂的特点,提出了一种基于特征重加权的遥感小样本目标检测算法RE-FSOD。该模型包括3部分:元特征提取器、特征重加权提取器、预测模块,其中元特征提取器由CSPDarknet-53、FPN以及PAN构成,负责提取数据的元特征；特征重加权提取器用于生成特征重加权向量,用于调整元特征来强化对于检测新类有帮助的特征；预测模块由YOLOv3的预测模块构成,在此基础上将定位损失函数替换为CIOU损失函数,提升模型的定位精度。最后在NWPU VHR-10遥感数据集上进行了训练和测试,实验结果表明,该方法相较于基线方法FSODM的在3-shot、5-shot、10-shot情况下分别提升了约19%、11%、8%。

Abstract:

A few-shot object detection model based on feature reweighting is proposed for remote sensing images with variable scale fuzzy target and complex background. The model consists of three parts: meta feature extractor, feature reweighting extractor and prediction module. The meta feature extractor is composed of CSPDarknet-53, Feature Pyramid Network and Path Aggregation Network, which is responsible for extracting meta features of data. The feature reweighting extractor is used to generate feature reweighting vectors, which are used to adjust meta features to enhance features that are helpful for detecting new classes. The prediction module is composed of the prediction module of YOLOv3. On this basis, the positioning loss function is replaced by CIOU to improve the positioning accuracy of the model. Finally, training and testing are carried out on the NWPU VH R-10 remote sensing data set. Finally, training and testing are carried out on the NWPU VHR-10 remote sensing data set. The experimental results show that compared with the baseline method FSODM, the improves by 19%, 11% and 8% respectively at 3-shot, 5-shot and 10-shot conditions.