| 摘 要: 针对遥感图像旋转框定位过程通常会出现分类分数和定位精度不匹配、交并比计算不精确的问题,文章提出了一个鲁棒性高的单阶段对齐检测网络(Single-stage Alignment Network,SAN),首先在网络中采用了对齐卷积,解决了分类分数和定位精度不匹配的问题;其次在网络训练过程中引入椭圆损失函数,将传统采用的定位交并比的计算方式转换为椭圆区域的参数表示计算。该方法在DOTA和HRSC2016数据集上进行的实验取得了74.3%和89.0%的平均精度,分别比基线方法高出13.2%和15.5%,优于大部分的主流网络模型。 |
| 关键词: 目标检测;遥感图像;卷积神经网络 |
|
中图分类号: TP311
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(61972187);福建省自然科学基金项目(2020J02024);福建省自然科学基金项目(2020J01828) |
|
| Remote Sensing Image Target Detection Algorithm Based on Alignment Convolution and Elliptic Loss Function |
|
ZHENG Lingyun1, LI Zuoyong2,3, CAI Yuanzheng2,3,4, TANG Zhengyi1
|
(1.College of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350118, China;
2.College of Computer and Control Engineering, Minjiang University, Fuzhou 350121, China;
3.Fujian Provincial Key Laboratory of In f ormation Processing and Intelligent Control, Fuzhou 350121, China;
4.Istrong Technology Corporation, Fuzhou 350100, China)
2211308055@smail.fjut.edu.cn; fzulzytdq@126.com; yuanzheng_cai@mju.edu.cn; tangzy84@126.com
|
| Abstract: Aiming at the common problems of mismatched classification scores and positioning accuracy, as well as inaccurate calculation of Intersection over Union (IoU) in the process of remote sensing image rotation frame positioning, this paper proposes a robust Single-stage Alignment Network (SAN). Firstly, alignment convolution is adopted in the network to solve the problem of mismatched classification scores and positioning accuracy. Secondly, in the process of network training, an elliptic loss function is introduced to convert the traditional calculation method of positioning IoU into parameter representation calculation in the elliptic region. This method achieves average accuracy of 74.3% and 89.0% respectively in experiments on the DOTA and HRSC2016 datasets, which is 13.2% and 15.5% higher than the baseline method, and outperforms most of the state-of-art network models. |
| Keywords: object detection; remote sensing images; convolutional neural network |
