作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2021, Vol. 47 ›› Issue (10): 186-193. doi: 10.19678/j.issn.1000-3428.0059234

• 图形图像处理 • 上一篇    下一篇

一种改进的BR-YOLOv3目标检测网络

宦海, 陈逸飞, 张琳, 李鹏程, 朱蓉蓉   

  1. 南京信息工程大学 电子与信息工程学院, 南京 210044
  • 收稿日期:2020-08-01 修回日期:2020-09-10 发布日期:2020-09-21
  • 作者简介:宦海(1978-),男,副教授,主研方向为通信与信息处理、图像处理;陈逸飞,硕士研究生;张琳,本科生;李鹏程、朱蓉蓉,硕士研究生。
  • 基金资助:
    国家自然科学基金(41671345)。

An Improved BR-YOLOv3 Object Detection Network

HUAN Hai, CHEN Yifei, ZHANG Lin, LI Pengcheng, ZHU Rongrong   

  1. School of Electronic and Information Engineering, Nanjing University of Information Science and Technology, Nanjing 210044, China
  • Received:2020-08-01 Revised:2020-09-10 Published:2020-09-21

摘要: 在目标检测任务中不同目标间尺寸差异较大,导致多尺寸目标难以被有效检测。基于YOLOv3提出BR-YOLOv3目标检测网络。利用空洞卷积提升网络层感受野尺寸的特性,使用不同数量、尺寸、膨胀率的卷积构建多层并行的空洞感受野模块。通过双向特征金字塔结构实现浅深层特征的双向融合,提升浅层预测分支分类、深层预测分支目标定位能力。使用LOSSGIOU定位损失函数实现目标回归过程整体化,从而降低目标漏检率。实验结果表明,BR-YOLOv3目标检测网络在Pascal VOC测试集上的测试平均精度均值达到79.24%,相比原网络提升3.52个百分点,且在检测精度上优于SSD、Faster RCNN等主流目标检测网络。

关键词: 目标检测, 目标尺寸差异, 空洞感受野模块, 双向特征金字塔, 定位损失函数

Abstract: In the object detection task, there is a large size difference between different objects, which makes it difficult to effectively detect object with multiple size.Based on YOLOv3, the Bidirectional FPN Atrous Reception YOLOv3 (BR-YOLOv3) target detection network is proposed.Using atrous convolution can effectively improve the receptive field size of the network layer, using different numbers, convolution kernel size, and dilation rate convolution to build a multi-layer parallel Atrous Receptive Module(ARM), and by using Bidirectional Feature Pyramid Structure Network(BiFPN) realizes bidirectional fusion of shallow and deep features, improving the classified ability of shallow prediction branch, and enhancing the ability of deep prediction branch's target positioning.By using the LOSSGIOU positioning loss function, the target regression process is integrated, and the target miss rate is reduced.Experimental results show that the improved RB-Yolov3 on the Pascal VOC test set has a mean average precision of 79.24%, which is an increase of 4.65% on the basis of the original network.It is superior to mainstream target detection networks such as SSD and Faster RCNN in detection accuracy.

Key words: object detection, object size difference, atrous receptive field module, Bidirectional Feature Pyramid Network (BiFPN), location loss function

中图分类号: