摘要
随着深度学习方法的快速发展,目标检测作为计算机视觉领域中最基本、最具有挑战性的任务之一,已取得了令人瞩目的进展。现有的算法大多针对于具有一定尺寸或比例的大中型目标,但由于待测目标尺寸小、特征弱等原因,对小目标的检测性能还远远不能令人满意。小目标检测(SOT)作为一种广泛应用于室外远程拍摄和航空遥感场景的技术,近年来受到了广泛的关注,各种方法层出不穷,但是目前对该问题的全面综述较少。从问题定义、算法分析、应用介绍、方向展望等方面对基于深度学习的小目标检测研究进展进行了综述。首先,给出了小目标检测问题的定义,阐述了其技术难点及在实际应用中面临的挑战;接着,从8个不同角度分析了检测器对小目标检测精度较低的主要原因及相应的改进方法,详细归纳总结了小目标检测在各技术方面的研究工作;然后介绍了几个特定场景下小目标检测算法的典型应用;最后,对小目标检测未来的发展趋势进行展望,提出可行的研究方向,期望为该领域的研究工作提供可借鉴和参考的思路。
Object detection,one of the most fundamental and challenging tasks in computer vision,has achieved remarkable breakthroughs with the rapid development of deep learning methods.However,most of the current state-of-the-art algorithms are designed for medium or large objects with regular sizes or proportions.The performance of detecting small objects is still far from satisfactory due to the small size and weak features of target objects.In recent years,Small Object Traching(SOT),which is widely applied in outdoor remote shooting and aerospace remote sensing scenarios,has drawn significant attention and substantial approaches have emerged.However,few comprehensive reviews on this issue have been conducted.This paper summarizes the research progress of deep learning-based SOT methods in terms of problem definition,algorithm analysis,application introduction,and future prospects.We first present the definition of the SOT problem and illustrate its technical difficulties and challenges faced in practical applications.The main reasons for the low accuracy of small object detectors and the corresponding improvement methods from eight different perspectives are then analyzed,and the SOT research work in various technical aspects summarized in detail.The representative applications of SOT algorithms in several specific scenarios are introduced.Finally,we prospect the development trends and promising directions for future research.Hopefully,this survey can provide reference for the research work in this field.
作者
李红光
于若男
丁文锐
LI Hongguang;YU Ruonan;DING Wenrui(Research Institute of Unmanned System,Beihang University,Beijing 100191.China;School of Electronics and Information Engineering,Beihang University,Beijing 100191.China)
出处
《航空学报》
EI
CAS
CSCD
北大核心
2021年第7期100-118,共19页
Acta Aeronautica et Astronautica Sinica
基金
装备预研领域基金(61403110404)
国家自然科学基金(62076019)。
关键词
目标检测
深度学习
小目标检测
计算机视觉
卷积神经网络
object detection
deep learning
small object traching
computer vision
convolutional neural networks
