电子学报 ›› 2022, Vol. 50 ›› Issue (2): 267-272.DOI: 10.12263/DZXB.20210354

• 学术论文 • 上一篇    下一篇

双向特征融合与特征选择的遥感影像目标检测

肖进胜1, 张舒豪1, 陈云华2, 王元方1, 杨力衡1   

  1. 1.武汉大学电子信息学院,湖北 武汉 430072
    2.广东工业大学计算机学院,广东 广州 510006
  • 收稿日期:2021-03-14 修回日期:2022-01-06 出版日期:2022-02-25 发布日期:2022-02-25
  • 作者简介:肖进胜 男,1975年7月出生,湖北武汉人.博士,武汉大学电子信息学院副教授,主要研究方向为视频图像处理、计算机视觉.E-mail: xiaojs@whu.edu.cn
    张舒豪 男,1995年9月出生,河南项城人.武汉大学电子信息学院硕士生,主要研究方向为遥感影像目标检测.E-mail: zhangsh@whu.edu.cn
    陈云华(通讯作者) 女,1977年3月出生,湖北仙桃人.博士,广东工业大学计算机学院副教授,主要研究方向为神经形态类脑计算、计算机视觉.E-mail: yhchen@gdut.edu.cn
  • 基金资助:
    广东省自然科学基金(2021A1515012233);中国科学院光电信息处理重点实验室开放课题基金(OEIP-O-202009)

Remote Sensing Image Object Detection Based on Bidirectional Feature Fusion and Feature Selection

XIAO Jin-sheng1, ZHANG Shu-hao1, CHEN Yun-hua2, WANG Yuan-fang1, YANG Li-heng1   

  1. 1.School of Electronic Information,Wuhan University,Wuhan,Hubei 430072,China
    2.School of Computer Science and Technology,Guangdong University of Technology,Guangzhou,Guangdong 510006,China
  • Received:2021-03-14 Revised:2022-01-06 Online:2022-02-25 Published:2022-02-25

摘要:

遥感影像中复杂的背景占据图像的大部分区域,严重影响了目标检测效果.本文提出一种可以对特征图进行多特征选择的目标检测网络.设计了双向多尺度特征融合网络,融合深浅层信息,提高复杂背景下小目标的检测效果,在保留常规特征金字塔自上而下路径的同时,增加一条自下而上的路径,减少浅层特征传递到顶层经历的网络层数,从而控制浅层特征损失.为了降低多尺度特征图中无用信息对后续检测网络的干扰,设计了基于注意力机制的多特征选择模块,网络自适应地专注于有用特征,忽略无用特征.针对传统五参数回归法在预测角度时存在严重的边界不连续问题,不能精确预测长宽比值比较大的目标,将角度预测当作分类任务处理.在DOTA数据集和自制数据集DOTA-GF上进行实验,6类典型目标的mAP分别达到0.651和0.641,与主流目标检测算法的对比实验结果表明提出的方法的有效性.

长摘要
遥感影像目标检测可广泛应用于环境管理、区域规划等诸多领域。由于目标具有多尺度性、方向任意性,背景复杂且占据较大区域,因此,遥感影像检测具有很大的挑战性。论文针对上述挑战性问题,提出了针对性的解决方案,并取得较好实验结果。包括:1)针对遥感影像目标多尺度的特性,设计了双向多尺度特征融合网络,在保留常规特征金字塔自上而下路径的同时,增加一条自下而上的路径,减少浅层特征传递到顶层过程中的信息损失,提升小目标的检测效果。2)针对遥感图像背景复杂的特点,设计了基于注意力机制的多特征选择模块,自适应地选择适合不同任务的特征,降低特征图中无用信息对检测结果的影响。3)针对传统五参数回归法在预测遥感影像目标角度时存在严重的边界不连续、不能精确预测长宽比较大的目标等问题,将角度预测当作分类任务处理。4)针对当前开源遥感影像数据集中的国产数据不足的问题,收集了部分GF-2和GF-6影像,自制了国产遥感影像数据集DOTA-GF。在DOTA和DOTA-GF数据集上,6类典型目标的mAP分别达到0.651和0.641,高于现有与主流目标检测算法。其中,本文方法在在船只、桥梁、小型汽车、存储罐这四类目标上都获得了最高的AP,在飞机和大型汽车这两类目标上的AP也接近最高值。

关键词: 遥感影像, 目标检测, 特征融合网络, 多特征选择, 角度预测

Abstract:

In remote sensing image object detection, the complex background always occupies a large area of the entire image, which seriously affects the object detection effect. This paper proposes an object detection network that can perform multiple feature fusion and selection on feature maps. A feature fusion network is used to fuse deep and shallow features to improve the detection effect of small objects in complex background. While retaining the up-bottom path of the feature fusion network, it adds a bottom-up path to diminish the number of network layers that the shallow features need to pass on to the top layer, thereby reducing the loss of shallow features. In order to reduce the interference of useless information in the fusion feature maps with detection network, a multiple feature selection module is designed. The attention mechanism in the multiple feature selection module enables the network to adaptively focus on more important features, ignore useless features. Since the conventional five-parameter regression method has serious boundary problems, the angle prediction is often inaccurate for objects with a large aspect ratio, to solve this problem, the proposed method treats angle prediction as a classification task. The mAP of our method on DOTA and self-made dataset DOTA-GF reaches 0.651 and 0.641, and the comparative experiments with mainstream object detection methods demonstrate the effectiveness of the proposed method.

Extended Abstract
Remote sensing image object detection can be widely used in many fields such as environmental management and regional planning. It is very challenging because objects in remote sensing images are multi-scale and have arbitrary orientations, and the background is complex and occupies a large area. This paper proposes targeted solutions to the above challenging problems, and achieves better experimental results. First, aiming at the multi-scale problem of the objects, we design a bidirectional multi-scale feature fusion network. While retaining the top-down path of the conventional feature pyramid, a bottom-up path is added to reduce information loss in the process of transferring from shallow features to the top layer, and to improve the detection effect of small targets. aiming at the complex background of remote sensing images, we design a multi-feature selection module based on the attention mechanism to adaptively select features suitable for different tasks and reduce the impact of useless information in the feature map on the detection results. Third, Aiming at the problem that the traditional five-parameter regression method suffers from serious boundary discontinuities in predicting the target angle of remote sensing images, and cannot accurately predict targets with large length and width ratios, we treat angle prediction as a classification task. In response to the problem of insufficient domestic data in the current open source remote sensing image datasets, we collected some GF-2 and GF-6 images and made our own domestic remote sensing image dataset DOTA-GF. On the DOTA and DOTA-GF datasets, the mAP of six kind of typical targets reaches 0.651 and 0.641, respectively, which are higher than the existing mainstream object detection algorithms. Specifically, the proposed method achieves the highest APs on all four classes of objects, including ships, bridges, small cars, and storage tanks, and the APs are also close to the highest values on two classes of objects, planes and large cars.

Key words: remote sensing image, object detection, feature fusion network, multiple feature selection, angle prediction

中图分类号: