计算机与现代化

• 图像处理 • 上一篇    下一篇

一种基于通道重排的轻量级目标检测网络

  

  1. (中国电子科技集团公司第十五研究所,北京100083)
  • 收稿日期:2019-06-28 出版日期:2020-03-03 发布日期:2020-03-03
  • 作者简介:徐晗智(1994-),男,辽宁葫芦岛人,硕士研究生,研究方向:机器学习和计算机视觉,E-mail: xhz1528@163.com; 艾中良(1971-),男,河北石家庄人,研究员,硕士,研究方向:网络安全,大数据,人工智能,E-mail: aizhongliang@hotmail.com; 张志超(1990-),男,河南信阳人,高级工程师,博士,研究方向:人工智能,计算机视觉,FPGA编程,E-mail: beyond9017@163.com。
  • 基金资助:
    国防科技创新特区基金资助项目(18-H863-01-ZT-005-009-01)

A Lightweight Target Detection Network Based on Channel Rearrangement 

  1. The 15th Research Institute of CETC, Beijing 100083, China)
  • Received:2019-06-28 Online:2020-03-03 Published:2020-03-03

摘要: Tiny YOLO和YOLOv3-tiny作为2种轻量级目标检测算法以其突出的速度表现而闻名。本文以这2种网络模型为基础,结合分组卷积并改进通道重排算法,改进了原来的损失函数,构建了一种新的更快的网络模型,通过改进YOLOv3的损失函数而增加其检测准确度。在PASCAL VOC数据集和COCO数据集上分别训练并且测试,该网络模型每秒处理的速度超过265张图片,Map值达到55.8%,准确度超过Tiny YOLO且与YOLOv3-tiny相仿。

关键词: 目标检测, 网络模型, 损失函数, 通道重排

Abstract: Tiny YOLO and YOLOv3-tiny are two lightweight target detection algorithms known for their outstanding speed performance. Based on these two network models, combining packet convolution and improved channel rearrangement algorithm and the original loss function, this paper constructs a new faster network model which improves the detection accuracy by improving the loss function of YOLOv3. The PASCAL VOC and COCO datasets were trained and tested respectively. The speed of the network model was faster than 265 pictures per second, and the accuracy was higher than Tiny YOLO and similar to YOLOv3-tiny.

Key words: target detection, network model, loss function, channel shuffle

中图分类号: