基于均匀设计的船舶目标检测深度学习 模型训练方法研究
DOI:
作者:
作者单位:

东北林业大学交通学院

作者简介:

通讯作者:

中图分类号:

U692.3

基金项目:

国家自然科学基金青年项目(51108137)


Research on Training Method of Deep Learning to Ship Target Detection Based on Uniform Design
Author:
Affiliation:

School of Traffic and Transportation, Northeast Forestry University

Fund Project:

Youth Project of National Natural Science Foundation of China(51108137)

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

    针对沿岸监控视频图像的舰船目标识别问题,传统检测方法检测速度和准确率有待进一步提升。深度学习方法已经能够实现船舶目标的检测,但模型网络结构复杂且存在交互影响的超参数,模型训练时耗长,检测精度往往受限于训练样本和超参数的选取。拟在YOLO算法的卷积神经网络模型训练过程中,采用均匀设计理念,对影响因素和水平协同优化,减少按照主观经验训练深度学习算法参数的盲目性。分析清晰度和人工标注影响,以场景、训练集与测试集样本比例、训练算法、训练轮次作为影响因素,设置差异化水平值,采用均匀设计方法训练深度学习模型,选用AP 值对比分析经验法和均匀设计法。以上海市黄浦江水域场景作为实例验证,研究表明:经验法设计的9组训练方案中最优AP值为0.84,均匀设计法产生的6组训练方案中,选用YOLOv5x算法,样本比例90%、训练100轮次时,深度学习模型AP值为0.91,表明均匀设计在深度学习模型超参数训练中有效。

    Abstract:

    In coastal surveillance video images of ship target recognition problem, the speed and accuracy of traditional detection methods need to be further improved. The deep learning method has been able to recognize ship targets, but the model network structure is complex and there are interaction effects of hyper parameters, the model training time is long, and the identification accuracy is often limited by the selection of training samples and hyper parameters. In the course of CNN model training of YOLO algorithm, the concept of uniform design is planned to be adopted to coordinate optimization of the level values of multiple influencing factors, so as to reduce the blindness of adjusting the parameters of deep learning algorithm according to subjective experience. After analyzing the impact of clarity and manual annotation, the scene, sample ratio of training set, training algorithm and training rounds were taken as the influencing factor. Differentiated level values were set and the deep learning model was trained by uniform design method. Using the value of AP to compare and analyze the empirical method and uniform design method. Taking three water areas of Huangpu River in Shanghai as example, the results show that: Among the 9 groups of schemes designed by the general empirical method, the optimal value of AP is 0.84. Among the 6 group of schemes generated by the uniform design method, when YOLOv5x algorithm is selected, the sample proportion is 90% and 100 training rounds, the accuracy of deep learning model is high, and the optimal AP value is 0.91. It is concluded that the effectiveness of the uniform design method in hyper parameter training of deep learning model.

    参考文献
    相似文献
    引证文献
引用本文

徐慧智. 基于均匀设计的船舶目标检测深度学习 模型训练方法研究[J]. 科学技术与工程, , ():

复制
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:2021-11-21
  • 最后修改日期:2022-04-24
  • 录用日期:2022-04-30
  • 在线发布日期:
  • 出版日期:
×
关于近期《科学技术与工程》编辑部居家办公的说明
亟待确认的版面费信息