《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (S2): 89-98.DOI: 10.11772/j.issn.1001-9081.2023010011

• 人工智能 • 上一篇    下一篇

改进循环生成对抗网络的车牌数据集自动生成算法

李文杰1, 张足生1(), 董阿妮2, 周坤晓1, 郭小红1   

  1. 1.东莞理工学院 网络空间安全学院,广东 东莞 523808
    2.东莞城市学院 计算机与信息学院,广东 东莞 523419
  • 收稿日期:2023-01-06 修回日期:2023-04-01 接受日期:2023-04-03 发布日期:2023-06-06 出版日期:2023-12-31
  • 通讯作者: 张足生
  • 作者简介:李文杰(1997—),男,山西大同人,硕士研究生,CCF会员,主要研究方向:计算机视觉;
    张足生(1980—),男,湖南衡阳人,教授,博士,CCF会员,主要研究方向:机器学习、无线传感器网络;
    董阿妮(1978—),女,陕西西安人,副教授,硕士,CCF会员,主要研究方向:无线传感器网络;
    周坤晓(1981—),男,湖北钟祥人,副教授,博士,主要研究方向:无线传感器网络;
    郭小红(1997—),女,江西九江人,硕士研究生,CCF会员,主要研究方向:机器学习。
  • 基金资助:
    国家自然科学基金资助项目(61872083);广东省自然科学基金资助项目(2019A1515011123);广东省普通高校重点领域专项(2020ZDZX3054)

Automatic generation algorithm of license plate dataset based on improved CycleGAN

Wenjie LI1, Zusheng ZHANG1(), Ani DONG2, Kunxiao ZHOU1, Xiaohong GUO1   

  1. 1.College of Cyberspace Security,Dongguan University of Technology,Dongguan Guangdong 523808,China
    2.College of Computer and Information,Dongguan City College,Dongguan Guangdong 523419,China
  • Received:2023-01-06 Revised:2023-04-01 Accepted:2023-04-03 Online:2023-06-06 Published:2023-12-31
  • Contact: Zusheng ZHANG

摘要:

针对现有车牌生成算法不能解决真实车牌数据集存在的数量少、多样性不足、字符标签不均衡、包含个人隐私等问题,提出一种改进循环生成对抗网络(CycleGAN)车牌生成算法。该算法由三部分组成:根据标准合成虚拟车牌,用仿射变换将虚拟车牌嵌入背景图像,然后由改进CycleGAN生成车牌图像样本。该算法通过引入权重解调机制解决了生成图像的白斑问题;利用重要区域损失、通道注意力与空间注意力实现了仅生成车牌部分的同时,保留了背景环境;采用最小二乘损失(LSLoss)改善了生成图像质量。已公开发布了20 000多张包括大倾角、远距离、模糊、复杂光照、天气条件等场景的生成车牌图像数据集,并通过对比实验验证了所提算法的有效性。在OpenITS、CLPD、CCPD(11K)验证集上的实验结果表明:在车牌检测任务中,与真实训练集的性能相近;在车牌识别任务中,识别精度相较于真实数据集分别提高了74.0%、28.0%、48.7%,相较于Bj?rklund、Duan、Han的算法都有3.0%以上提高。所提算法可生成数量多、多样性高、字符标签均衡、无隐私问题的车牌数据,能够对车牌检测与识别算法的训练提供有效支持。

关键词: 生成对抗网络, 数据生成, 车牌识别, 深度学习, 智慧交通

Abstract:

Aiming at the existing license plate generation algorithms that cannot solve the problems of low quantity, inadequate diversity, unbalanced character labels, and personal privacy in real license plate datasets, an improved CycleGAN (Cycle-consistent Generative Adversarial Network) license plate generation algorithm was proposed. The algorithm was consisted of three parts: synthesized license plates were made up according to corresponding standards, affine transformation was used to embed synthesized plates into background images, and improved CycleGAN was used to generate license plate samples. Weight Demodulation (WD) was introduced to solve white spot problem of generated images; Significant Area Loss (SALoss), Channel Attention Mechanism (CAM) and Spatial Attention Mechanism (SAM) were used to generate license plate parts while preserving background environment; and Least Square Loss (LSLoss) was utilized to improve the quality of generated images. A generated license plate image dataset with over 20 000 images was released including scenarios such as large inclinations, long distances, blur, complex lighting, and weather conditions, and the effectiveness of the proposed algorithm was verified in comparison experiments. Experimental results on OpenITS, CLPD, and CCPD (11K) validation sets show that, for the license plate detection task, the performance was similar to that on real training sets; for the license plate recognition task, the accuracies were improved by 74.0%, 28.0%, and 48.7% respectively compared to that on real training set, while over 3.0% improvement compared to Bj?rklund's, Duan's, and Han's algorithms. The proposed algorithm can generate license plate data with large quantities, high diversity, balanced character labels, and no privacy issues, which can effectively support the training of license plate detection and recognition tasks.

Key words: Generative Adversarial Network (GAN), data generation, license plate recognition, deep learning, intelligent transportation

中图分类号: