1 南京信息工程大学江苏省大气环境与装备技术协同创新中心，江苏 南京 210044
2 南京信息工程大学自动化学院，江苏 南京 210044
针对遥感有向目标存在的检测问题，设计了一个基于改进Rotated RPN的网络，设计特征重组机制，通过加权使网络关注有效目标区域。使用新的有向框标注方法，避免在临界角度出现错位等问题。在检测头前端使用极化注意力模块，改善因为分类和回归任务所需特征不一致导致的性能下降问题。实验结果表明，该模型可以提高多类目标的检测精度。相较于基准Rotated RPN，该模型在Dior-R数据集上精度提升4.95%，在HRSC2016数据集上精度提升11.75%。
遥感 有向目标检测 深度学习 特征重组 极化注意力
Author Affiliations
Abstract
1 Postdoctoral Mobile Research Station of the School of Electronic Engineering, Heilongjiang University, Harbin 150080, China
2 School of Applied Sciences, Harbin University of Science and Technology, Harbin 150080, China
3 Key Laboratory of Electronics Engineering, College of Heilongjiang Province, Heilongjiang University, Harbin 150080, China
We experimentally investigate the effects of the surface roughness of gold thin films on the properties of surface plasmon resonance. By annealing at different temperatures, film samples with different surface morphologies are obtained. Specifically, due to the diffusion of the gold atoms towards the films’ surface, the surface root-mean-square roughness decreases with the increasing annealing temperature. Then, we measure the surface plasmon resonance of the samples. The results show that the resonance angle of the surface plasmon resonance is sensitive to the root-mean-square roughness, and it gradually decreases by reducing the surface root-mean-square roughness.
240.6680 Surface plasmons 310.6860 Thin films, optical properties Chinese Optics Letters
2016, 14(4): 042401
