计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (11): 16-27.DOI: 10.3778/j.issn.1002-8331.2211-0377

• 热点与综述 • 上一篇    下一篇

深度学习小目标检测算法综述

董刚,谢维成,黄小龙,乔逸天,毛骞   

  1. 西华大学 电气与电子信息学院,成都 610039
  • 出版日期:2023-06-01 发布日期:2023-06-01

Review of Small Object Detection Algorithms Based on Deep Learning

DONG Gang, XIE Weicheng, HUANG Xiaolong, QIAO Yitian, MAO Qian   

  1. School of Electrical Engineering and Electronic Information, Xihua University, Chengdu 610039, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 现有的目标检测算法,对大目标以及中目标的检测已具有较高的准确率,然而由于小目标在图像中的像素以及可利用的特征较少等原因,导致小目标的检测精度相较于大目标而言过低。通过融合特征层,小目标的检测已取得了不错的效果,但仍存在对于微小目标的定位等问题。基于此,解释了小目标的定义,指出了导致小目标检测精度低的五点原因。将近几年最新进展以及过往经典的小目标检测优化方法按照大致原理从多尺度特征、评估指标、超分辨率等方面进行叙述。归纳了针对特定场景下的小目标检测:航空遥感图像以及人脸行人的检测方法。总结并提出了未来小目标检测可能的研究方向。

关键词: 小目标, 目标检测, 计算机视觉, 深度学习

Abstract: The existing object detection algorithms have high accuracy for the detection of large objects and medium objects, but due to the few pixels in the image and the available features of small objects, the detection accuracy of small objects is too low compared with that of large objects. By fusing the feature layer, the detection of small objects has achieved good results, but there are still problems such as the localization of small objects. Based on this, the definition of small objects is first explained, and five reasons for the low detection accuracy of small objects are pointed out. Subsequently, the latest progress in recent years and the classic small object detection optimization method in the past are described from multi-scale features, novel metric, and super-resolution according to the general principle. Secondly, the detection methods of small objects for specific scenes:aerial images, faces, and pedestrians are summarized. Finally, the possible research directions of small object detection in the future are summarized and proposed.

Key words: small object, object detection, computer vision, deep learning