YOLOv5预测边界框分簇自适应损失权重改进模型
作者:
作者单位:

江西财经大学软件与物联网工程学院

作者简介:

通讯作者:

中图分类号:

TP301

基金项目:

国家自然科学基金项目(61866014)


Enhanced Self-adaptive Loss Weight YOLOv5 Model Based on Predicted Bounding Boxes in Clusters
Author:
Affiliation:

School of Software & Internet of Things Engineering, Jiangxi University of Finance and Economics

Fund Project:

The National Natural Science Foundation of China (61866014)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    目标检测的精确程度是计算机视觉识别任务的主要影响因素。针对单阶段目标检测模型 YOLOv5 存在的检测精度问题,从多任务损失优化角度,提出一种在不同分辨率特征图上基于同一目标的预测边界框分簇自适应损失权重改进模型。该模型由 GT(Ground True)目标边界框 UID 分配器、GT 目标边界框 UID 匹配器、边界框位置与分类损失权重算法构成,通过改善 YOLOv5 的位置精度和分类精度实现模型整体精度的提升。实验结果表明,改进模型平均精度均值(Mean Average Precision,mAP)较 YOLOv5.6 标准模型平均相对提升了5.23%;相较更为复杂的 YOLOv5x6 标准模型,改进模型 mAP 取得了 8.02% 的相对提升。

    Abstract:

    The object detection precision plays a critical role in computer vision task. Aiming at the precision problem available in the one-stage object detection model of YOLOv5, this paper proposes an enhanced self-adaptive loss weight YOLOv5 model based on predicted bounding boxes in clusters presenting the individual targets in multi-resolution feature maps to optimize the multi-task loss. The enhanced model consists of GT(Ground True) target bounding box UID distributor, GT target bounding box UID matcher, bounding box position loss weight algorithm and classification loss weight algorithm. The overall detection precision is improved by the enhancements of both position precision and classification precision in YOLOv5. The experimental results present that compared with YOLOv5.6, the mean average precision(mAP) is promoted relatively by 5.23% on average by the enhanced model which achieves the relative performance of 8.02% compared with the more complex model of YOLOv5x6.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2021-09-15
  • 最后修改日期:2021-12-31
  • 录用日期:2022-01-11
  • 在线发布日期: 2022-02-01
  • 出版日期: