Abstract:

To solve the problems of large scale difference and low detection accuracy in remote sensing image object detection task, an improved YOLOv3 model for remote sensing image object detection based on weighted strategy is proposed. Firstly, the detection branch of feature map with smaller receptive fields is added to improve small-size object detection accuracy in remote sensing image. Secondly, anadaptive weighted fusion method for multi-scale feature maps is designed, which improves object detection accuracy by excavating representation capability of feature extraction network and fusing the multi-scale features comprehensively. Finally, a new remote sensing image object detection dataset containing four classes of objects in DIOR dataset is built, using which the proposed model is trained and tested. Experimental results show that the improved model achieves mean Average Precision（mAP） of 80.25%, which is 8.2% higher than that using the original YOLOv3 model. The trained model is also tested on RSOD, UCAS-AOD and NWPU VHR-10 datasets, which verifies the adaptability of the improved model.

Key words: remote sensing image, target detection, multi-scale features fusion, weighting strategy

摘要：

针对遥感图像目标检测任务中存在的目标尺度差异大、检测精度低等问题，提出了一种基于加权策略的改进YOLOv3遥感图像目标检测模型。为提高对遥感图像中小目标的检测精度，增加具有较小感受野的特征图像的检测分支。设计了一种多尺度特征图像自适应加权融合方法，通过挖掘特征提取网络的表征能力，综合利用多尺度特征提高了目标检测精度。采用DIOR数据集的4类目标构建了一个新的遥感图像目标检测数据集，并进行了改进模型的训练与测试。实验结果表明，改进后的模型取得了80.25%的平均精度均值（mean Average Precision，mAP），相比于改进前提高了8.2%。将训练模型对RSOD、UCAS-AOD、NWPU VHR-10数据集进行测试，验证了改进模型具有较好的适应性。

关键词: 遥感图像, 目标检测, 多尺度特征融合, 加权策略