计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (5): 176-193.DOI: 10.3778/j.issn.1002-8331.2207-0389

• 图形图像处理 • 上一篇    下一篇

YOLOv5定位多特征融合的车标识别

董光辉,陈星宇   

  1. 东北林业大学 信息与计算机工程学院,哈尔滨 150040
  • 出版日期:2023-03-01 发布日期:2023-03-01

Vehicle Logo Recognition with YOLOv5 Location and Multi-Feature Fusion

DONG Guanghui, CHEN Xingyu   

  1. College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Online:2023-03-01 Published:2023-03-01

摘要: 为解决智能交通系统中车标识别的问题,提出YOLOv5s网络车标定位多特征融合的车标图像识别方案。车标定位阶段选择YOLOv5s网络以满足对车标定位速度与精度等的需求。车标识别阶段通过调整扩展高斯差分中的参数得到具有不同效果的车标边缘,设计一组二维Gabor滤波器对边缘检测后的车标图像进行滤波处理并提取出对应的车标图像特征向量,通过计算待测车标图像特征与标准比对库中特征向量的欧几里德距离,取距离最小者对应的标签索引作为分类识别结果,该方案的最佳识别正确率为96.91%。采用随机森林算法进行分类后的最佳识别正确率可达99.33%。该方案的车标定位与识别最佳整体正确率超过了YOLOv5s网络直接一步到位识别车标的方案,且相较于传统图像处理方法有明显提升。

关键词: 车标识别, YOLOv5s, 多特征融合, 扩展高斯差分, 二维Gabor滤波, 欧几里德距离, 随机森林

Abstract: In order to solve the problem of vehicle logo identification in intelligent transportation system, a vehicle logo image recognition scheme based on using YOLOv5s network to locate vehicle logo and multi-feature fusion is proposed. The YOLOv5s network is selected in the vehicle logo positioning stage to meet the requirements of the positioning speed and accuracy. Firstly, in the vehicle logo recognition stage, the vehicle logo edges with different effects are obtained by adjusting the parameters in the extended difference of Gaussians. Then, a set of two-dimensional Gabor filters are designed to filter the edge-detected vehicle logo image and extract the corresponding feature vector. Finally, by calculating the Euclidean distance between the feature vector of the vehicle logo image to be tested and the feature vector in the standard comparison library, the label index corresponding to the smallest distance is taken as the classification and recognition result. The optimal recognition accuracy of the scheme is 96.91%, which can reach 99.33% after classification with random forest algorithm. The best overall accuracy of vehicle logo positioning and recognition in this paper exceeds that by using the YOLOv5s network in one step, and it is significantly improved compared with the traditional image processing method.

Key words: vehicle logo identification, YOLOv5s, multi-feature fusion, extended difference of Gaussians, two dimensional Gabor filtering, Euclidean distance, random forest