• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于YOLOv3算法的训练集优化和检测方法的研究

高星1,刘剑飞1,郝禄国2,董琪琪1   

  1. (1.河北工业大学电子信息工程学院,天津 300401;2.广东工业大学信息工程学院,广东 广州 510006)
  • 收稿日期:2019-05-27 修回日期:2019-08-05 出版日期:2020-01-25 发布日期:2020-01-25
  • 基金资助:

    天津市自然科学基金(15JCYBJC17000);河北省高等学校科学技术研究重点项目(ZD2017021)

A training set optimization and detection
method based on YOLOv3 algorithm

GAO Xing1,LIU Jian-fei1,HAO Lu-guo2,DONG Qi-qi1   

  1.  (1.School of Electronic Information Engineering,Hebei University of Technology,Tianjin 300401;
    2.School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)
     
  • Received:2019-05-27 Revised:2019-08-05 Online:2020-01-25 Published:2020-01-25

摘要:

YOLOv3是一种单步目标检测算法,不需要产生区域候选网络(RPN)来提取目标信息,相对于双步目标检测算法具有更快的检测速度。但是,现有算法在小目标检测上存在精度不高和漏检现象的问题,为此提出了一种基于YOLOv3算法的训练集优化和图层处理的检测方法。首先在标准数据集VOC2007+2012和自建的举手行为数据集上采用K-means算法做聚类分析,以得到适应数据集训练尺寸的anchor大小;然后通过调整训练参数及选择合理的标签标注方式进行训练;最后对输入图像进行图层处理并进行目标检测。实验结果表明,聚类分析后VOC2007验证集的平均准确度(mAP)提高了14%,并有效解决了原算法在检测过程中较高卷积层上感受野小的问题,从而使YOLOv3算法在小目标物体的检测上精度提高,漏检率也相对下降。
 

关键词: YOLOv3, anchor, 小目标检测, 聚类分析

Abstract:

YOLOv3 is a one-stage target detection algorithm that does not generate Region Proposal Network (RPN) to extract target information. Compared with the two-stage target detection algorithm, it has faster detection speed. However, the existing algorithms has the problems of low accuracy and missed detection in small target detection. Therefore, based on YOLOv3 algorithm, a detection method of training set optimization and layer processing is proposed. Firstly, K-means algorithm is used to cluster the standard dataset VOC2007+2012 and the self-built behavior dataset, so as to get the anchor size that fits the training size of the dataset. Then, The training is carried out by adjusting the training parameters and selecting a reasonable labeling method. Finally, the input image is processed by layer and the target is detected. The experimental results show that the average accuracy (mAP) of VOC2007 verification set is improved by 1.4% after clustering analysis, the problem of small feeling field on the higher convolution layer is effectively solved during the detection process of the original algorithm, and the accuracy of YOLOv3 algorithm is improved when detecting small target objects and reducing the missed detection rate.
 

Key words: YOLOv3, anchor, small target detection, cluster analysis