计算机与现代化 ›› 2020, Vol. 0 ›› Issue (09): 83-88.doi: 10.3969/j.issn.1006-2475.2020.09.015

• 算法设计与分析 • 上一篇    下一篇

轻量级YOLOv3的交通标志检测算法

  

  1. (1.长春工业大学计算机科学与工程学院,吉林长春130012;2.吉林大学软件学院,吉林长春130012)
  • 收稿日期:2020-02-03 出版日期:2020-09-24 发布日期:2020-09-24
  • 作者简介:白士磊(1992—),男,河北邯郸人,硕士研究生,研究方向:目标检测,图像处理,E-mail: 2201703042@stu.ccut.edu.cn; 殷柯欣 (1975—),女,教授,博士,研究方向:模式识别,网络与信息安全,E-mail: yinkexin@126.com。
  • 基金资助:
    国家自然科学基金资助项目(61572229); 吉林省教育厅项目(20180401)

Lightweight YOLOv3 Traffic Sign Detection Algorithm

  1. (1. School of Computer Science & Engineering, Changchun University of Technology, Changchun 130012, China;
    2. College of Software, Jilin University, Changchun 130012, China)
  • Received:2020-02-03 Online:2020-09-24 Published:2020-09-24

摘要: 交通标志检测在自动驾驶领域一直是个比较热门的课题。在深度学习算法中,YOLOv3和Faster R-CNN已经获得了极好的目标检测性能,但在检测小目标时,存在漏检的情况。针对交通标志检测中小目标准确快速识别的需求,本文提出一种轻量级YOLOv3的交通标志检测算法。通过卷积神经网络同时使用浅层和深层的特征提取,得到多尺度特征图,深层特征可以有效地保持检测精度不下降,浅层特征可以有效地提高小目标检测任务的精度。通过剪枝算法对模型进行压缩,将训练好的模型进行稀疏训练,把一些不重要的卷积核通道删除掉,对剪枝后的模型微调,保持模型文件中参数的平衡,同时保持检测精度。实验结果表明,通过提取多尺度特征图的方法模型准确率提高了2.3%,通过剪枝算法对模型压缩,使模型的权重大小减小了70%,模型的检测时间节省了90%。由此建立了鲁棒性更强的轻量级交通标志检测模型,可以部署在移动端嵌入式设备上,不再占用庞大的GPU计算资源即可提高检测效率。

关键词: 卷积神经网络; 交通标志, 小目标检测; 多尺度特征图; 模型压缩

Abstract: Traffic sign detection has always been a hot topic in the field of autonomous driving. In deep learning algorithms, YOLOv3 and Faster R-CNN have obtained excellent object detection performance, but when detecting small targets, there are cases of missed detection. In order to accurately and quickly identify small targets in traffic sign detection, this paper proposes a lightweight YOLOv3 traffic sign detection algorithm. Multi-scale feature maps are obtained through convolutional neural networks while using both shallow and deep feature extraction. Deep features can effectively keep the detection accuracy from falling, and shallow features can effectively improve the accuracy of small target detection tasks. The model is compressed by the pruning algorithm, the trained model is sparsely trained, some unimportant convolution kernel channels are deleted, and the pruned model is fine-tuning to maintain the balance of parameters in the model file. The experimental results show that the accuracy of the model is improved by 2.3% by extracting the multi-scale feature map, the weight of the model is reduced by 70% by compressing the model with the pruning algorithm, and the detection time of the model is saved by 90%. Therefore, a lightweight traffic sign detection model with stronger robustness can be deployed on mobile embedded devices, without taking up large GPU computing resources but improving detection efficiency.

Key words: convolutional neural network, traffic signs, small object detection, multi-scale feature map, model compression

中图分类号: