测绘学报 ›› 2018, Vol. 47 ›› Issue (6): 873-881.doi: 10.11947/j.AGCS.2018.20170633

• 数字摄影测量与深度学习方法 • 上一篇    下一篇

深度残差网络的多光谱遥感图像显著目标检测

戴玉超1, 张静1,2, Fatih PORIKLI2, 何明一1   

  1. 1. 西北工业大学电子信息学院, 陕西 西安 710129;
    2. 澳大利亚国立大学工程研究院, 澳大利亚 堪培拉 2601
  • 收稿日期:2017-12-09 修回日期:2018-03-30 出版日期:2018-06-20 发布日期:2018-06-21
  • 通讯作者: 何明一 E-mail:myhe@nwpu.edu.cn
  • 作者简介:戴玉超(1982-),男,博士,教授,研究方向为计算机视觉与模式识别。E-mail:daiyuchao@nwpu.edu.cn
  • 基金资助:
    国家自然科学基金(61420106007;61671387);澳大利亚研究理事会DECRA项目(DE140100180)

Salient Object Detection from Multi-spectral Remote Sensing Images with Deep Residual Network

DAI Yuchao1, ZHANG Jing1,2, Fatih PORIKLI2, HE Mingyi1   

  1. 1. School of Electronics and Information, Northwestern Polytechnical University, Xi'an 710129, China;
    2. Research School of Engineering, Australian National University, Canberra 2601, Australia
  • Received:2017-12-09 Revised:2018-03-30 Online:2018-06-20 Published:2018-06-21
  • Supported by:
    The National Natural Science Foundation of China (Nos. 61420106007;61671387);The DECRA of Australian Research Council (No. DE140100180)

摘要: 本文侧重于介绍智能化摄影测量深度学习的深度残差方法。显著目标检测致力于自动检测和定位图像中人最感兴趣的目标区域。多波段遥感图像因其更加丰富的光谱信息和揭示观测目标物理属性的能力在目标检测中获得重要应用。传统的显著目标检测方法通过手工设计特征,计算图像各像素或者超像素与邻域像素或者超像素之间的对比度检测显著目标。随着深度学习的巨大发展,特别是全卷积神经网络的引入,基于深度卷积网络的显著目标检测算法取得重要进步。然而,由于数据获取和标记的困难,多波段遥感图像显著目标检测的研究依然主要采用手工设计特征。本文研究基于深度卷积神经网络的多波段遥感图像显著目标检测算法,提出一种基于深度残差网络的自上而下的多光波段遥感图像显著目标检测网络,该网络可以有效挖掘深度残差网络不同层次上的显著性特征,以端对端方式实现显著目标检测。为了应对多波段遥感图像数据量有限、无法训练深度残差网络的问题,本文提出通过浅层神经网络从RGB图像直接生成多波段遥感图像,实现光谱方向的超分辨率。在现有多波段遥感图像和可见光图像显著目标检测数据集上的试验结果超过当前最好方法10%以上,验证了本文方法的有效性。

关键词: 深度残差网络, 显著目标检测, 光谱超分辨率, 自上而下模型, 遥感图像处理

Abstract: This paper focuses on intelligent photogrammetry deep learning:deep residual method.Salient object detection aims at identifying the visually interesting object regions that are consistent with human perception.Multispectral remote sensing images provide rich radiometric information in revealing the physical properties of the observed objects,therefore promise a great potential in salient object detection tasks.Conventional salient object detection methods often employ handcrafted features to predict saliency by evaluating the pixel-wise or superpixel-wise similarity.With the recent emergence of deep learning based approaches,in particular,fully convolutional neural networks,there has been profound progress in visual saliency detection.However,this success has not been extended to multispectral remote sensing images,and existing multispectral salient object detection methods are still mainly based on handcrafted features,essentially due to the difficulties in image acquisition and labeling.In this paper,we propose a novel deep residual network based on a top-down model,which is trained in an end-to-end manner to tackle the above issues in multispectral salient object detection.Our model effectively exploits the saliency cues at different levels of the deep residual network.To overcome the limited availability of remote sensing images in training of our deep residual network,we also introduce a new spectral image reconstruction model that can generate multispectral images from RGB images.Our extensive experimental evaluations using both multispectral and RGB salient object detection datasets demonstrate a significant performance improvement of more than 10% compared with the state-of-the-art methods.

Key words: deep residual network, salient object detection, spectral super-resolution, top-down model, remote sensing image processing

中图分类号: