| 基于深度学习的空间非合作目标特征检测与识别 |
| |
| 引用本文： | 李林泽,张涛.基于深度学习的空间非合作目标特征检测与识别[J].智能系统学报,2020,15(6):1154-1162. |
| |
| 作者姓名： | 李林泽 张涛 |
| |
| 作者单位： | 清华大学 自动化系, 北京 100084 |
| |
| 摘 要： | 针对空间非合作目标检测与识别任务的智能化要求，本文将深度学习方法Mask R-CNN(mask region-based convolutional neural network)应用于任务中，并借鉴R-FCN(region-based fully convolutional networks)和Light-head R-CNN (light-head region-based convolutional neural network)对其进行优化改进，提升检测速度，以满足空间任务实时性要求。实验结果表明，与传统的Mask R-CNN相比，改进的Mask R-CNN可缩短20%的检测时间。针对深度神经网络需要大样本数据集进行训练的特点，本文基于迁移学习提出搭建虚拟环境进行样本采集，构造空间目标特征检测与识别数据集的方法。实验结果表明，网络在虚拟环境生成的数据集上可以很好地学习到相应特征，从而具备迁移到实际任务的能力。
|
| 关 键 词： | 空间非合作目标 特征检测与识别 深度学习 区域全卷积网络 头部轻量化卷积神经网络 改进的Mask R-CNN 数据集构建 迁移学习 |
Feature detection and recognition of spatial noncooperative objects based on deep learning |
| |
| LI Linze,ZHANG Tao.Feature detection and recognition of spatial noncooperative objects based on deep learning[J].CAAL Transactions on Intelligent Systems,2020,15(6):1154-1162. |
| |
| Authors: | LI Linze ZHANG Tao |
| |
| Affiliation: | Department of Automation, Tsinghua University, Beijing 100084, China |
| |
| Abstract: | To meet the intelligence requirements of feature detection and recognition of a spatial noncooperative target, the deep learning method—Mask Region-based Convolutional Network (Mask R-CNN)—is applied to the task using the idea of Region-based Fully Convolutional Network (R-FCN) and Light-head R-CNN as references to improve the detection speed to meet the real-time requirements of spatial tasks. Results obtained from the aforementioned study shows that the modified Mask R-CNN can shorten the detection time by 20% compared with the original version. Considering that deep neural networks require large sample datasets for training, this paper proposes a method for constructing a virtual environment for sample collection to build datasets, thus constructing the feature detection of spatial objects and datasets for recognition. The experimental result shows that the network learns the corresponding characteristics very well on the database generated from the virtual environment, and thus, the network can be migrated to the real task. |
| |
| Keywords: | spatial noncooperative targets feature detection and recognition deep learning R-FCN light-head R-CNN modified mask R-CNN dataset construction transfer learning |
|
| 点击此处可从《智能系统学报》浏览原始摘要信息 |
|
点击此处可从《智能系统学报》下载全文 |
