计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (17): 187-194.DOI: 10.3778/j.issn.1002-8331.2205-0543

• 图形图像处理 • 上一篇    下一篇

基于混合类别均衡损失的车型精细识别

李熙莹,全峰玮,叶芝桧   

  1. 1.中山大学 智能工程学院 智能交通研究中心,广州 510006
    2.广东省智能交通系统重点实验室,广州 510006
    3.视频图像智能分析与应用技术公安部重点实验室,广州 510006
  • 出版日期:2023-09-01 发布日期:2023-09-01

Fine-Grained Vehicle Model Recognition Based on Mixed Class Balance Loss

LI Xiying, QUAN Fengwei, YE Zhihui   

  1. 1.Intelligent Transportation Research Center, School of Intelligent Systems Engineering, Sun Yat-sen University, Guangzhou 510006, China
    2.Guangdong Province Key Laboratory of Intelligent Transportation Systems, Guangzhou 510006, China
    3.Key Laboratory of Video and Image Intelligent Analysis and Application Technology, Guangzhou 510006, China
  • Online:2023-09-01 Published:2023-09-01

摘要: 为了应对车型精细识别中数据分布不均衡导致训练中头部类别过拟合,而尾部类别被忽略的问题,提出了一种基于混合类别均衡损失的车型精细识别数据增强方法。结合Mixup数据增强方法和类别均衡损失,提出混合类别均衡交叉熵损失函数;通过均衡子集微调的训练策略,进一步提高了长尾分布数据的识别效果。实验结果表明,算法在Stanford Cars、CompCars、SYSU Cars数据集上的识别准确率分别比Baseline提高了1.07、0.17和1.58个百分点,有效地缓解了因车型数据不均衡带来的问题,进一步提高了车型精细识别的识别效果。其中SYSU Cars为自建数据集,由66?137张车辆正脸图片构成,包含102种品牌,691种车型以及不同的光照条件(即将在OpenITS上公开)。

关键词: 车型精细识别, 细粒度识别, 混合类别均衡损失, 长尾分布

Abstract: In order to deal with the problem that the head category is overfitting and the tail category is ignored during training due to the uneven distribution of vehicle model data, a method for fine-grained vehicle model recognition based on mixed class balance loss is proposed. A mixed class balance cross-entropy loss function is proposed by combining the Mixup data augmentation method and the class-balance loss. The balanced subset fine-tuning is used as the training strategies to further improve the recognition effect of long-tailed distribution data. The experimental results show that the accuracy on Stanford Cars, CompCars, and SYSU Cars datasets is improved by 1.07, 0.17, and 1.58?percentage points, respectively, which effectively alleviates the problems caused by the imbalanced data and improves the recognition effect of vehicle model recognition even further. The SYSU Cars is a self-built dataset, which contains 102 vehicle brands, 691 models and various lighting scenes(to be available on OpenITS soon).

Key words: fine-grained vehicle model recognition, fine-grained recognition, mixed class balance loss, long-tailed distribution