基于YOLOv3的定向目标检测算法
Directional Target Detection Algorithm Based on YOLOv3
辛月兰 1朱杰 2谢琪琦2
作者信息
- 1. 青海师范大学物理与电子信息工程学院,青海 西宁 810008;省部共建藏语智能信息处理及应用国家重点实验室,青海 西宁 810001
- 2. 青海师范大学物理与电子信息工程学院,青海 西宁 810008
- 折叠
摘要
为解决YOLOv3 目标检测算法中无法对旋转物体进行定向目标检测的问题,提出一种基于YOLOv3 的定向目标检测算法.首先,使用多维坐标对训练集的图像进行定向标定,以适应网络训练;其次使用最小外接矩形对网络输出的矩形框进行修正优化,以获得更加准确贴合的检测框;然后对网络的损失函数进行改进,使其适应多维坐标的回归;最后,对改进后的网络进行训练.在UCAS-AOD数据集上的实验结果表明,目标检测的能力在改进后有了明显提升,比原始YOLOv3 算法精确率提高了 6.1%,召回率提高了 3.2%.
Abstract
In order to solve the problem that the YOLOv3 target detection algorithm cannot detect the directional target of rotating objects,this paper proposes a directional target detection algorithm based on YOLOv3.Firstly,this method used the multi-dimensional coordinate to align the training set to suit the training of the network.Scoendly,the rectangular box output from the network was optimized using the minimum outer rectangle to obtain a more accu-rate fit of the detection box.Next,the loss function of the network was improved to adapt it to the regression of multi-dimensional coordinates.Finally,the improved network wais trained.The experimental results on the UCAS-AOD dataset show that the ability of target detection is significantly improved after the improvement,with a 6.1%increase in accuracy and a 3.2%increase in recall over the original YOLOv3 algorithm.
关键词
定向目标检测/多维坐标/最小外接矩形Key words
Directional target detection/Multi-dimensional coordinates/Minimum external rectangle引用本文复制引用
基金项目
国家自然科学基金(61662062)
青海省自然科学基金面上项目(2022-ZJ-929)
出版年
2024
NETL
NSTL
万方数据