改进RRPN模型的遥感图像目标检测  

Remote sensing image object detection based on improved RRPN model

作　　者：鲁晓波[1] 郭艳光 辛春花 LU Xiaobo;GUO Yanguang;XIN Chunhua(Department of Computer Technology and Information Management,Inner Mongolia Agricultural University,Baotou 010010,China)

机构地区：[1]内蒙古农业大学计算机技术与信息管理系,内蒙古包头010010

出　　处：《现代电子技术》2025年第1期8-16,共9页Modern Electronics Technique

基　　金：国家自然科学基金项目(31960361)。

摘　　要：针对遥感目标背景复杂、易受外界环境干扰,传统方法无法满足复杂场景下的检测高精度与实时性要求的问题,提出基于改进RRPN模型的遥感图像目标检测方法。首先,将特征金字塔(FPN)架构引入到了模型的残差网络中,使得遥感图像的高、低层特征得到了有效融合;其次,在特征提取网络中添加了通道和空间相融合的注意力机制(CBAM),提升了模型在遥感图像目标特征提取方面的跨通道和空间处理能力;此外,将剔除重叠建议框时的原始NMS算法优化为DIoUNMS算法,综合考虑遥感图像候选框之间的重叠度、距离、尺度大小等因素,使目标框的回归过程更加稳定。对比实验与消融实验显示,所提方法在公共数据集DOTA和HRSC2016上获得的平均精度均值mAP分别可高达77.30%、90.24%,较原始RRPN模型分别提高了8.29%、11.16%,且优于其他几种较新的经典模型,表明所提方法对于复杂环境下的遥感图像目标检测是合理且有效的。The background of the remote sensing object is complex.In addition,the remote sensing object is susceptible to external environment,so the traditional methods fail to meet the requirements of high precision and real-time detection in complex scenes.In view of this,the paper proposes a remote sensing image object detection method based on the improved RRPN model.The framework of feature pyramid network(FPN)is introduced into the residual network of the model,which enabled the effective fusion of high-and low-level features of remote sensing images.The convolutional block attention mechanism(CBAM)combining channel and space is incorporated into the feature extraction network,so as to improve the cross-channel and spatial processing capability of the model in the feature extraction of remote sensing image object.In addition,the original NMS(non-maximum suppression)algorithm is optimized into DIoU-NMS algorithm for eliminating overlapping object frames,and the overlap,distance,scale and other factors among the candidate frames of remote sensing images are taken into account comprehensively,so as to make the regression of object frames more stable.In the comparative and ablation experiments,it is shown that the proposed method achieves mAP(mean average precision)of 77.30%and 90.24%on the public datasets DOTA and HRSC2016,respectively,which are 8.29%and 11.16%higher than that of the original RRPN(rotation region proposal network)model,and it is better than that of the other advanced classical models.This indicates that the proposed method is reasonable and effective for the object detection of remote sensing images in complex environments.

关 键 词：目标检测 遥感图像 带旋转的候选框算法 卷积通道注意力模块 DIoU-NMS 特征金字塔 DOTA HRSC2016数据集 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]