跨尺度移位的有向目标检测方法

Oriented Object Detection Method Based on Cross-Scale Shift

目标类内尺度差异大以及类间相似程度高给传统遥感目标检测带来挑战,多尺度信息融合以及基于FasterR-CNN的有向检测方法是解决尺度差异性及类间相似性的有效手段。但多尺度权重融合策略忽视了跨尺度对图像语义特征的提取导致检测精度低,同时基于FasterR-CNN的有向检测方法精度低速度慢。针对上述问题,提出一种跨尺度移位的有向目标检测方法。首先通过改进特征金字塔网络(FPN)实现多层特征图高效融合;其次在网络中加入跨尺度移位模块(CSM)提升FPN中多尺度特征间的相关性;最后采用有向区域建议网络(ORPN)提升了有向候选框的转换效率。本文方法分别在DOTA-v1.5和HRSC2016遥感数据集上开展测试,相比对照组,均值准确率(mAP)分别提升了3.68%和1.32%;同时在单块2080ti上用1024×1024图像进行测试,检测速度提升5.9%。

The large intra class scale differences and high inter class similarity of the target pose challenges to tra⁃ditional remote sensing target detection.Multi scale information fusion and directed detection methods based on Faste⁃rR-CNN are effective means to address scale differences and inter class similarity.However,the multi-scale weight fusion strategy neglects the extraction of image semantic features across scales,resulting in low detection accuracy.At the same time,the directed detection method based on FasterR-CNN has low accuracy and slow speed.In response to the above issues,a cross-scale shifted oriented object detection approach was put out.First,the feature pyramid net⁃work(FPN)was enhancedto fuse the multilayer feature map effectively.Then,the cross-scale shift module(CSM)was added to the network in order to improve the correlation between the multi-scale features in the FPN.Finally,the oriented region proposal network(ORPN)was employed to increase the conversion efficiency of the oriented bounding boxes.The DOTA-v1.5 and HRSC2016 remote sensing datasets were used to validate the methodology described in this work.Compared to the control group,the mean average precision(mAP)increases by 3.68 and 1.32 percent,respectively.The detection speed increases by 5.9%when performed on a single 2080ti with 1024×1024 images.