一种改进的遥感图像多目标识别算法  

An Improved Algorithm for Multi-object Recognition in Remote Sensing Images

作　　者：黎林 姚颖 郭鑫轶 叶云龙 LI Lin;YAO Ying;GUO Xinyi;YE Yunlong(School of Computer Science,Hubei Univ.of Tech.,Wuhan 430068,China)

机构地区：[1]湖北工业大学计算机学院,湖北武汉430068

出　　处：《湖北工业大学学报》2025年第2期42-48,共7页Journal of Hubei University of Technology

摘　　要：由于遥感图像背景复杂、目标尺度差异大,易导致多目标识别效果不佳,针对此问题提出了一种改进的遥感图像多目标识别算法。以Faster R-CNN网络为基础,采用EfficientNet作为主干网络以提取更完整的图像特征;然后引入改进后的特征金字塔结构,获得更多上下文信息;其次,统计数据集中所有目标的尺度信息,重新设置RPN网络锚框;接着,使用焦点损失函数解决正负样本不平衡的问题;最后利用ROI Align区域聚集解决区域不匹配问题。进行消融实验验证了该算法的可行性与有效性,实验结果表明该算法优于原始Faster R-CNN算法,多类平均准确率mAP达到了93.11%,提升了10.59%。Aiming at the problems of complex background and large difference of target scale in the process of remote sensing image object recognition,this paper proposes an improved remote sensing image multi-object recognition algorithm.Based on Fast R-CNN network,this paper uses EfficientNet as the backbone network to extract more complete image features;Then the improved feature pyramid structure is introduced to obtain more context information;Secondly,the scale information of all objects in the data set is counted,and the RPN network anchor frame is reset;Then,the focus loss function is used to solve the problem of imbalance between positive and negative samples;Finally,ROI align region aggregation is used to solve the region mismatch problem.In this paper,ablation experiments are carried out to verify the feasibility and effectiveness of the algorithm.The experimental results show that the algorithm in this paper is superior to the original Fast R-CNN algorithm,and the average multi class accuracy(mAP)reaches 93.11%,increasing by 10.59%.

关 键 词：遥感图像 Faster R-CNN 多目标检测 特征金字塔 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]