作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (2): 173-179. doi: 10.19678/j.issn.1000-3428.0059979

• 图形图像处理 • 上一篇    下一篇

融合改进FPN与关联网络的Faster R-CNN目标检测

汪常建, 丁勇, 卢盼成   

  1. 南京航空航天大学 自动化学院, 南京 210016
  • 收稿日期:2020-11-12 修回日期:2020-12-25 发布日期:2021-02-05
  • 作者简介:汪常建(1997-),男,硕士研究生,主研方向为目标检测与跟踪;丁勇,副教授;卢盼成,硕士研究生。
  • 基金资助:
    国家自然科学基金面上项目(61473146)。

Object Detection Using Faster R-CNN Combining Improved FPN and Relation Network

WANG Changjian, DING Yong, LU Pancheng   

  1. College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
  • Received:2020-11-12 Revised:2020-12-25 Published:2021-02-05

摘要: 在无人机场景下,目标检测存在样本数量不足、成像视角不同的问题,导致检测精度低。提出一种结合改进特征金字塔网络(FPN)与关联网络的Faster R-CNN目标检测算法。通过在传统FPN结构中以自下而上的特征融合方式提取特征图的语义信息和位置信息,最大程度地保留特征图的多尺度信息。同时利用候选区域之间的形状特征和位置特征构造区域之间的关联特征,并与深度特征相融合进行分类回归,从而充分提取特征图的整体信息,实现目标检测。在PASCAL VOC 2007和NWPU VHR-10数据集上的实验结果表明,相比FPN+Faster R-CNN算法,该算法的交并比和平均检测精度分别提高了10和2.7个百分点,具有较优的目标检测性能。

关键词: 目标检测, 尺度, 特征金字塔网络, 关联网络, 特征融合

Abstract: The existing target detection usually suffers from insufficient sample number and different image angles, resulting in low detection accuracy. This paper proposes a target detection algorithm using Faster R-CNN by combining improved Feature Pyramid Network(FPN) structure with a relation network.Base on the traditional FPN structure, the bottom-up feature fusion process is added to extract rich semantic information and location information of the feature map of targets.Then the location information and shape features between candidate regions are used to construct relation features, which are subsequently fused with deep features, so fully extract the overall information of the feature map and realize target detection.The experimental results on PASCAL VOC 2007 and NWPU VHR-10 datasets show that compared with the FPN+Faster R-CNN algorithm, the proposed algorithm increases the Intersection over Union (IoU) by about 10 percentage points and detection accuracy by about 2.7 percentage points, displaying excellent performance in target detection.

Key words: object detection, scale, Feature Pyramid Network(FPN), relation network, feature fusion

中图分类号: