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 ›› 2022, Vol. 48 ›› Issue (9): 269-276,285. doi: 10.19678/j.issn.1000-3428.0062671

• 开发研究与工程应用 • 上一篇    下一篇

基于弱语义分割的轻量化交通标志检测网络

曾雷鸣1,3, 侯进2,3, 陈子锐2,3, 周浩然1,3   

  1. 1. 西南交通大学 计算机与人工智能学院, 成都 611756;
    2. 西南交通大学 信息科学与技术学院, 成都 611756;
    3. 西南交通大学 综合交通大数据应用技术国家工程实验室, 成都 611756
  • 收稿日期:2021-09-13 修回日期:2021-10-26 发布日期:2021-11-02
  • 作者简介:曾雷鸣(1992—),男,硕士研究生,主研方向为深度学习、目标检测;侯进(通信作者),副教授、博士;陈子锐、周浩然,硕士研究生。
  • 基金资助:
    四川省科技计划项目(2020SYSY0016)。

Lightweight Traffic Sign Detection Network Based on Weak Semantic Segmentation

ZENG Leiming1,3, HOU Jin2,3, CHEN Zirui2,3, ZHOU Haoran1,3   

  1. 1. School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China;
    2. School of Information Science and Technology, Southwest Jiaotong University, Chengdu 611756, China;
    3. National Engineering Laboratory of Comprehensive Transportation Big Data Application Technology, Southwest Jiaotong University, Chengdu 611756, China
  • Received:2021-09-13 Revised:2021-10-26 Published:2021-11-02

摘要: 针对现有网络在检测高分辨率交通标志图片时速度过慢、精确度较低等问题,提出一种轻量化交通标志检测网络。在MobileNetv3-Large基础上对YOLOv4网络的骨干部分进行优化,针对数据集的特点舍弃部分耗时层,更改第8层和第14层的输出通道数,并改进基础模块中通道域注意力网络的注意力机制,使输出的权重数值能更准确地表征特征的重要程度。在检测头前加入基于弱语义分割的动态增强附件,利用其输出作为空间权重分布来矫正激活区域,以避免提取能力下降导致误检、漏检问题,最终构成YOLOv4-SLite网络。采用滑窗剪裁的方法对高分辨率图片进行训练和预测,从而减少训练时间及增加样本的多样性。在TT100K交通标志数据集上的实验结果表明,相较于YOLOv4基准网络,YOLOv4-SLite网络的mAP@0.5仅下降了0.2%,但模型大小减少了96.5%,响应速度提升了227%,精确度与速度的平衡效果达到了预期。

关键词: 交通标志检测, YOLOv4网络, 轻量化网络, 弱语义分割, 注意力机制

Abstract: Aiming at the problems of slow speed and low accuracy in detecting high-resolution traffic sign images in existing networks, a lightweight traffic sign-detection network is proposed.On the basis of MobileNetv3-Large, this study optimizes the backbone of a YOLOv4 network, discards some time-consuming layers according to the characteristics of the dataset, changes the number of output channels of layers 8 and 14, and improves the attention mechanism of Squeeze and Excitation Network (SENet) in the basic module, so that the weight value of the output can more accurately represent the importance of the characteristics.This study adds a dynamic enhanced attachment based on weak semantic segmentation in front of the detection header, and uses its output as the spatial weight distribution to correct the active region, to avoid the problem of false detection and missed detection caused by the decline of extraction ability, and finally form a YOLOv4-SLite network.The sliding window clipping method is used to train and predict high-resolution images, to reduce the training time and increase the diversity of samples.The experimental results on the TT100K traffic sign dataset show that, compared with the YOLOv4 benchmark network, the mAP@0.5 of the YOLOv4-SLite network is lost by 0.2%, but the model size is reduced by 96.5%, and the response speed is increased by 227%.The balance of accuracy and speed achieved meets the expectation.

Key words: traffic sign detection, YOLOv4 network, lightweight network, weak semantic segmentation, attention mechanism

中图分类号: