电子学报 ›› 2022, Vol. 50 ›› Issue (7): 1684-1695.DOI: 10.12263/DZXB.20200917

• 学术论文 • 上一篇    

基于特征融合和自学习锚框的高分辨率图像小目标检测算法

李超, 黄新宇, 王凯   

  1. 湖北工业大学计算机学院,湖北 武汉 430068
  • 收稿日期:2020-08-21 修回日期:2021-11-05 出版日期:2022-07-25 发布日期:2022-07-30
  • 作者简介:李 超 男,1982年出生于湖北省洪湖市.湖北工业大学计算机学院硕士生导师.研究方向为计算机视觉、机器学习.E-mail: lich.mail@163.com
    黄新宇 男,1998年出生于湖北省黄石市.湖北工业大学计算机学院本科生.研究方向为目标检测与语义分割.E-mail: 864546664@qq.com
  • 基金资助:
    湖北省自然科学基金(2017CFB326);国家级大学生创新创业训练计划项目(201810500016)

Small Object Detection of High-Resolution Images Based on Feature Fusion and Learnable Anchor

LI Chao, HUANG Xin-yu, WANG Kai   

  1. School of Computer,Hubei University of Technology,Wuhan,Hubei 430068,China
  • Received:2020-08-21 Revised:2021-11-05 Online:2022-07-25 Published:2022-07-30

摘要:

为了提高高分辨率图像中小目标的检测精度,解决高分辨率图像在下采样和局部裁切时由于细节和背景信息丢失造成的漏检和误检问题,本文提出了一种基于特征融合和自学习锚框的小目标检测算法.算法采用多路分支网络对高分辨率图像的全局语义和细节特征平滑后逐层融合,以同时增强特征图上小目标的细节和背景特征.针对训练样本尺寸差异造成不同分支网络上特征表达不一致的问题,本文引入自学习锚框使融合后的特征图能够适应锚框的位置和形状.使用本文算法与目前先进的目标检测算法对下采样图像和切块检测,大量实验结果验证了本文算法对高分辨率图像小目标检测的准确性和有效性.

关键词: 小目标检测, 特征融合, 自学习锚框, 高分辨率图像

Abstract:

Small object detection of high-resolution images presents significant challenges. To solve the problem that downsampling and cropping of high-resolution images result in missed detections and false detections due to the loss of fine details and contextual information, an algorithm based on feature fusion and learnable anchor is proposed for small object detection of high-resolution images. Contextual and detailed features are extracted from downsampled images and cropped patches respectively, which are then fused layer-wise. The fused features are further combined with smoothed features to strengthen both fine details and contextual information. To mitigate the feature inconsistency, learnable anchor is applied to make the fused features accommodative to the location and shape of anchors. The proposed method is tested from the perspective of global inference and local inference compared to state-of-the-art detectors. The experimental results show the accuracy and effectiveness of the proposed method.

Key words: small object detection, feature fusion, learnable anchor, high resolution images

中图分类号: