摘要
针对高分辨率遥感图像具有物体尺度差异大、小目标排列密集且方向性强的问题,提出一种基于多尺度特征提取的旋转遥感目标检测算法。选用CenterNet作为基准模型,对其进行重新设计。首先,为增强上下文信息提取能力,结合多尺度空洞卷积提出并使用感受野扩展模块;其次,结合自适应特征融合,增强算法对多尺度目标的提取能力;最后,重新设计CenterNet检测头并更新损失函数,增强模型对旋转物体的检测性能。设计的模型命名为CenterNet for remote sensing images(CenterNet-RS)。在DOTA数据集上进行实验,CenterNet-RS的平均精度均值(mAP)达73.01%,相较于基准模型,提高了9.45个百分点。实验结果表明,设计的方法可有效提高算法检测遥感图像目标的精度。
A rotation remote sensing target detection algorithm based on multi-scale feature extraction is proposed,because high-resolution remote sensing images have large object scale differences,dense small-object arrangements,and strong orientation.In this study,CenterNet was chosen as the benchmark model and redesigned.First,to improve the context information extraction ability,we proposed and applied the receptive field expansion module combined with multiscale cavity convolution.Second,the extraction ability of the algorithm for multi-scale targets was improved in combination with adaptive feature fusion.Finally,we redesigned the CenterNet detection head and updated the loss function to improve the detection performance of the model for rotating objects.The designed model is named CenterNet for remote sensing images(CenterNet-RS).Experiments were performed on the DOTA dataset,and the mean average precision(mAP)of CenterNet-RS reaches 73.01%,which is 9.45 percentage points higher than the baseline model.Thus,the experimental findings demonstrate that the proposed method can significantly increase the target detection accuracy for remote sensing images.
作者
吴洛冰
谷玉海
吴文昊
范帅鑫
Wu Luobing;Gu Yuhai;Wu Wenhao;Fan Shuaixin(Key Laboratory of Modern Measurement and Control Technology,Ministry of Education,Beijing Information Science&Technology University,Beijing 100089,China;Mechanical Electrical Engineering School,Beijing Information Science&Technology University,Beijing 100089,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第12期443-451,共9页
Laser & Optoelectronics Progress
基金
国家自然科学基金(51975058)
学科建设专项资助项目(5112011015)。
关键词
遥感图像
旋转目标检测
多尺度
卷积神经网络
特征融合
remote sensing image
rotating object detection
multi-scale
convolutional neural network
feature fusion
