东北大学学报:自然科学版 ›› 2019, Vol. 40 ›› Issue (2): 174-179.DOI: 10.12068/j.issn.1005-3026.2019.02.005

• 信息与控制 • 上一篇    下一篇

基于深度神经网络剪枝的两阶段遥感图像目标检测

王生生, 王萌, 王光耀   

  1. (吉林大学 计算机科学与技术学院, 吉林 长春130012)
  • 收稿日期:2017-11-30 修回日期:2017-11-30 出版日期:2019-02-15 发布日期:2019-02-12
  • 通讯作者: 王生生
  • 作者简介:王生生(1974-),男,吉林长春人,吉林大学教授,博士生导师.
  • 基金资助:
    国家自然科学基金资助项目(61472161); 吉林省科技发展计划项目(20180101334JC,20190302117GX).

Deep Neural Network Pruning Based Two-Stage Remote Sensing Image Object Detection

WANG Sheng-sheng, WANG Meng, WANG Guang-yao   

  1. College of Computer Science and Technology, Jilin University, Changchun 130012, China.
  • Received:2017-11-30 Revised:2017-11-30 Online:2019-02-15 Published:2019-02-12
  • Contact: WANG Meng
  • About author:-
  • Supported by:
    -

摘要: 在高分辨率遥感图像目标检测中,受云雾、光照、复杂背景、噪声等因素影响,现有目标检测方法虚警率高、速度慢、精确度低.为此提出基于深度神经网络剪枝的两阶段目标检测(object detection based on deep pruning,ODDP)方法.首先,给出深度神经网络剪枝方法,基于深度神经网络剪枝分别提出自主学习区域提取网络算法与优化训练分类网络算法;然后,将上述两算法用于卷积神经网络,得到两阶段目标检测模型.实验结果表明,在NWPU VHR-10高分辨率遥感数据集上,相比现有目标检测方法,ODDP的检测速度和精度均有一定提升.

关键词: 计算机视觉, 目标检测, 高分辨率遥感图像, 深度学习, 卷积神经网络

Abstract: In the object detection of high-resolution remote-sensing images, affected by cloud, light, complex background, noise and other factors, the existing object detection method has high false alarm, low speed and low precision. So we propose a two-stage object detection method based on deep pruning. First, we propose deep pruning, and then based on the deep pruning we propose an algorithm that learns region proposal network automatically and an algorithm that we train classification network with optimizing training method. We then apply the two algorithms to convolutional neural network and get a two-stage object detection model. The experiment result shows that our method has a certain improvement on precision and speed compared with the state-of-the-art method.

Key words: computer vision, object detection, high-resolution remote sensing image, deep learning, convolutional neural network

中图分类号: