基于深度学习的遥感目标检测技术

Remote sensing object detection based on deep learning

针对遥感目标检测精度不足,提出一种改进YOLOv5s遥感目标检测算法。主干网络采用CSP-D模块进行特征提取,充分利用深层和浅层特征进行特征增强;颈部网络采用BiFPN结构进行特征融合,提高多尺度特征信息融合效率。实验结果表明,针对遥感目标数据集DIOR,改进YOLOv5s网络平均准确率均值(mAP)提升2.1%,不同目标类别平均准确率(AP)均有提升,缓解原网络检测存在的漏检误检问题,改进网络检测速度仍能满足实时性要求,具有更优的检测性能。

Aiming at the lack of detection accuracy about remote sensing target detection,an improved YOLOv5s remote sensing target detection algorithm was proposed to improve the detection accuracy of remote sensing targets.CSP-D module was used for feature extraction in backbone network,and the feature information of the shallow and deep layers of the network was fully utilized for feature enhancement.The neck network used BiFPN structure for feature fusion to improve the efficiency of multi-scale feature information fusion.Experimental results show that for the remote sensing target dataset DIOR,compared with the original network,the mean average precision(mAP)is increased by 2.1%through the improved YOLOv5s network.The average precision(AP)of different object detection classes is improved.The problems of missed detection and false detection exist in the original network detection are alleviated.The improved network detection speed can still meet the real-time requirements and shows better detection performance.