电子学报 ›› 2020, Vol. 48 ›› Issue (2): 390-397.DOI: 10.3969/j.issn.0372-2112.2020.02.023

• 学术论文 • 上一篇    下一篇

融合深度扩张网络和轻量化网络的目标检测模型

权宇1, 李志欣1, 张灿龙1, 马慧芳1,2   

  1. 1. 广西师范大学广西多源信息挖掘与安全重点实验室, 广西桂林 541004;
    2. 西北师范大学计算机科学与工程学院, 甘肃兰州 730070
  • 收稿日期:2019-04-22 修回日期:2019-10-17 出版日期:2020-02-25
    • 通讯作者:
    • 李志欣
    • 作者简介:
    • 权宇 女,1992年9月出生,江苏徐州人.广西师范大学计算机科学与信息工程学院硕士研究生.研究方向为图像理解与机器学习.E-mail:quanyu0919@163.com;张灿龙 男,1975年10月出生,湖南娄底人.现为广西师范大学计算机科学与信息工程学院教授、博士生导师.研究领域为目标跟踪与模式识别.E-mail:zcltyp@163.com;马慧芳 女,1981年7月出生,甘肃兰州人.博士,硕士生导师,现为西北师范大学计算机科学与工程学院教授.研究领域为数据挖掘与机器学习.E-mail:mahuifang@yeah.net
    • 基金资助:
    • 国家自然科学基金 (No.61966004,No.61663004,No.61762078,No.61866004); 广西自然科学基金 (No.2019GXNSFDA245018,No.2016GXNSFAA380146,No.2017 GXNSFAA198365,No.2018GXNSFDA281009); 广西多源信息挖掘与安全重点实验室基金 (No.16-A-03-02,No.MIMS18-08)

Fusing Deep Dilated Convolutions Network and Light-Weight Network for Object Detection

QUAN Yu1, LI Zhi-xin1, ZHANG Can-long1, MA Hui-fang1,2   

  1. 1. Guangxi Key Lab of Multi-source Information Mining and Security, Guangxi Normal University, Guilin, Guangxi 541004, China;
    2. College of Computer Science and Engineering, Northwest Normal University, Lanzhou, Gansu 730070, China
  • Received:2019-04-22 Revised:2019-10-17 Online:2020-02-25 Published:2020-02-25
    • Corresponding author:
    • LI Zhi-xin
    • Supported by:
    • National Natural Science Foundation of China (No.61966004, No.61663004, No.61762078, No.61866004); Natural Science Foundation of Guangxi Zhuang Autonomous Region,  China (No.2019GXNSFDA245018, No.2016GXNSFAA380146, No.2017GXNSFAA198365, No.2018GXNSFDA281009); Guangxi Key Laboratory Foundation for Multi-source Information Mining and Security (No.16-A-03-02, No.MIMS18-08)

摘要: 目标检测作为计算机视觉的一个重要研究方向,近年来在算法性能上有了突破性进展.为了更好的提升两阶段目标检测的精度与速度性能,提出了一种基于迁移学习方法的融合深度扩张卷积网络和轻量化网络的检测模型.首先用扩张卷积网络替换主干网络中部分的卷积残差模块——深度扩张卷积网络D_dNet-65;然后对预训练后的特征图进行压缩操作,并增加一个81类的全连接层以确保正常进行分类和回归操作——轻量化网络结构;最后,引入迁移学习方法并融合D_dNet和轻量化网络结构,通过迁移实现模型的进一步优化.实验在典型的数据集MSCOCO以及VOC07上进行.实验评估表明,本文提出的方法具有良好的有效性和可扩展性.

关键词: 图像目标检测, 迁移学习, 扩张卷积网络, 轻量化网络, 卷积神经网络

Abstract: Object detection is an important research direction in the field of computer vision.In recent years, object detection has made great advances in public datasets, and there are also breakthroughs in algorithmic performance. In order to improve the accuracy and speed performance of two-stage object detection, this paper proposes a detection model based on transfer learning method that fuses the deep dilated convolutions network and the light-weight network. First, the dilated convolutions network is used to replace the convolutional residual module in the backbone network, namely deep dilated convolution network (D_dNet-65). Then, by compressing the pretrained feature map and adding an 81-class fully connected layer to replace the original two layers, namely light-weight network. Finally, the transfer learning method is introduced in the pretraining to optimize the model (D_dNet and light-weight network). The experiment was carried out on a typical data set, MSCOCO and VOC07. And the experiment shows that the method proposed in this paper has good effectiveness and scalability.

Key words: image object detection, transfer learning, dilated convolution network, light-weight network, convolution neural network

中图分类号: