东北大学学报(自然科学版) ›› 2021, Vol. 42 ›› Issue (5): 609-615.DOI: 10.12068/j.issn.1005-3026.2021.05.001

• 信息与控制 •    下一篇

YOLOv3-ADS:一种基于YOLOv3的深度学习目标检测压缩模型

宋欣1,2, 李奇1, 解婉君1, 李宁3   

  1. (1. 东北大学 计算机科学与工程学院, 辽宁 沈阳110169; 2. 中国科学院 自动化研究所, 北京100190; 3. 中国移动信息技术有限公司, 北京100037)
  • 修回日期:2020-07-20 接受日期:2020-07-20 发布日期:2021-05-20
  • 通讯作者: 宋欣
  • 作者简介:宋欣(1978-),女,黑龙江齐齐哈尔人,东北大学副教授.
  • 基金资助:
    基金项目;(半空) 基金项目.国家自然科学基金资助项目(61603083); 中央高校基本科研业务费专项资金资助项目(N162304009); 河北省高等学校科学技术研究重点项目(ZD2017303); 中国科学院自动化研究所国家重点实验室开放课题项目(20180105).

YOLOv3-ADS:A Compression Model for Deep Learning Object Detection Based on YOLOv3

SONG Xin1,2, LI Qi1, XIE Wan-jun1, LI Ning3   

  1. 1. School of Computer Science & Engineering, Northeastern University, Shenyang 110169, China; 2. Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China; 3.China Mobile Information Technology Co., Ltd., Beijing 100037, China.
  • Revised:2020-07-20 Accepted:2020-07-20 Published:2021-05-20
  • Contact: SONG Xin
  • About author:-
  • Supported by:
    -

摘要: 基于原有YOLOv3模型占用存储空间较大,所需初始化数据集样本和参数较多的问题,本文提出了一种基于YOLOv3的深度学习目标检测压缩模型YOLOv3-ADS.该模型使用拼接、叠加等方法对较少的有代表性的初始数据集进行数据增强,引入了DIoU损失函数,提升了目标检测的准确度.最后,通过稀疏训练和剪枝率阈值设置实现了YOLOv3-ADS模型的压缩处理,减少了模型实现过程中的冗余节点、参数数量和所需存储空间.实验结果表明,提出的YOLOv3-ADS压缩模型与已有的YOLOv3模型相比,平均精度值(mAP值)提升了约30%,由0.6418提升至0.8368,需设置参数量下降了96.6%,由原来的63.0MB降至2.2MB,在保证了较高目标检测准确率的同时,YOLOv3-ADS模型所需存储空间下降了96.5%,由252MB降至仅需8.81MB.

关键词: 目标检测;YOLOv3-ADS模型;深度学习;YOLOv3模型;压缩模型

Abstract: Since the original YOLOv3 model takes up a large amount of storage space, more initialization dataset samples and parameters are required. A deep learning object detection compression model YOLOv3-ADS was proposed based on YOLOv3. The proposed model uses the methods of splicing and stacking for data enhancement of the fewer representative initial datasets. It introduces the DIoU loss function, and improves the accuracy of object detection. Finally, YOLOv3-ADS model was compressed by sparse training and pruning rate threshold setting, which reduces the number of redundant nodes, parameters and the required storage space. The experimental results show that, compared with the existing YOLOv3 model, the average accuracy of the proposed YOLOv3-ADS compression model is increased by about 30%, from 0.6418 to 0.8368, and the number of parameters is reduced by 96.6%, from the original 63.0MB to 2.2MB. At the same time, the storage space required by YOLOv3-ADS model is reduced by 96.5%, from 252MB to only 8.81MB.

Key words: object detection; YOLOv3-ADS model;deep learning;YOLOv3 model;compression model

中图分类号: