R-P-Faster R-CNN day65 读论文:高分辨率遥感影像综合地理空间目标检测框架
An Efficient and Robust Integrated Geospatial Object Detection Framework for High Spatial Resolution Remote Sensing Imagery
- 1. Introduction
- 3. Overview of the Proposed R-P-Faster R-CNN Framework
- 3.1. 有效集成区域建议网络与目标检测Faster R-CNN框架
- 3.1.2. RPN与FASTER-R-CNN共享卷积特征的集成策略
- 7. Conclusions
1. Introduction
提出了一种基于FASTER区域卷积神经网络(FASTER R-CNN)的高效、鲁棒的综合地理空间目标检测框架。
- 这篇文章有不少之前已经学过或了解了的基础知识,所以咱们直接跳过繁杂的实验和介绍,直捣黄龙看看框架是什么样的,记住Keywords:High Spatial Resolution、 Remote Sensing Image、FASTER R-CNN
- 这篇文章对Faster R-CNN有一个较为详细的解释
3. Overview of the Proposed R-P-Faster R-CNN Framework
所提出的R-P-Faster R-CNN框架由三个主要过程组成,即:
-
有效的Faster R-CNN过程
-
鲁棒有效的预训练过程以弥补标记训练样本的不足
-
以及有效的时间守恒过程
3.1. 有效集成区域建议网络与目标检测Faster R-CNN框架
Faster R-CNN通过共享卷积特性,将RPN和Faster R-CNN融合在一起,以交替训练的方式,利用多任务损失函数对整个网络进行优化。 这个过程描述如下。
Faster R-CNN的框架
锚点的选择原则
多任务损失函数
ZF模型和VGG模型的结构:(a)ZF模型的结构; 和(b)VGG模型的结构
3.1.2. RPN与FASTER-R-CNN共享卷积特征的集成策略
提出的R-P-Faster R-CNN框架的流程图
7. Conclusions
In this paper, an effective R-P-Faster R-CNN object detection framework has been proposed forHSR remote sensing imagery.Considering the complex distribution of geospatial objects and thelow efficiency of the current object detection methods for HSR remote sensing imagery, the robustproperties of a transfer mechanism and the sharable properties of Faster R-CNN are considered andcombined in the R-P-Faster R-CNN object detection framework. The transfer mechanism provides thedeep learning based object detection algorithm with good initial network parameters. The sharableproperties of the R-P-Faster R-CNN object detection framework help save the time consumption ofre-training the neural network. The combination of these properties results in the proposec
本文提出了一种适用于HSR遥感图像的R-P快速R-CNN目标检测框架。考虑到当前HSR遥感图像目标检测方法的复杂分布和低效率,在R-P快速R-CNN目标检测框架中考虑了传输机制的鲁棒性和更快R-CNN的共享特性。该传输机制为基于深度学习的目标检测算法提供了良好的初始网络参数。R-P-更快的R-CNN目标检测框架的优点有助于节省神经网络的训练时间。这些属性的组合将导致提议
R-P-FasterR-P-更快R-CNN algorithm performing better than the current object detection methods.R-CNN算法比现有的目标检测方法性能更好。
In contrast to the CNN-based HSR imagery object detection methods,the region proposalgeneration stage and the object recognition stage are separated, which improves the time consumptionfor efficiently training the object detection framework. The feature-sharing mechanism of the proposedR-P-Faster R-CNN framework can effectively solve this problem and improve the performance ofHSR imagery geospatial object detection.The experimental results obtained with the NWPU VHR-10geospatial object detection dataset confirm that the proposed R-P-Faster R-CNN framework is efficientand effective. In our future work, a more effective object recognition framework will be considered forHSR imagery geospatial object detection.
与基于CNN的HSR图像目标检测方法相比,将区域提出阶段和目标识别阶段分离,提高了训练目标检测框架的时间消耗。本文提出的R-P-更快R-CNN框架的特征共享机制能够有效地解决这一问题,提高了HSR图像地理空间目标检测的性能。在NWPU VHR-10地理空间目标检测数据集上的实验结果表明,所提出的R-P-更快的R-CNN框架是有效和有效的。在我们未来的工作中,将考虑一个更有效的目标识别框架,用于HSR图像的地理空间目标检测。