摘要
针对道路交通环境中路面标志识别涉及的数据集较少和识别准确率不足的问题,研究了基于深度卷积生成对抗网络的道路表面指示标志的识别方法。在深度卷积生成对抗网络的结构基础上,根据具体应用修改生成网络和判别网络的损失函数,并用随机梯度下降算法替代原始的优化器,对指示标志的原始样本集进行样本生成,以增加样本数据量。基于Faster R-CNN算法进行路面标志的特征提取,实现路面指示标志的识别,并基于迁移学习对识别模型进行微调,将目标识别效果应用于实际道路环境中。实验结果表明,通过深度卷积生成对抗网络生成的样本图像有效地扩增了路面标志的数据集,增广后的多类目标识别的mAP提高了17.1%,小样本情况下的识别准确率随着样本量的增加和样本质量的改善而得到了明显的提高。
In studies of identification for road surface indicator signs,less datasets and insufficient recognition accuracy are major difficulties. A method based on the Deep Convolutional Generative Adversarial Networks(DCGAN) is developed to identify of road surface indication signs. Based on the structure of the deep convolutional generation of the adversarial network,the loss functions for generating and discriminating the network are modified according to the specific application,and the loss functions are used to generate and discriminate the network. The original optimizer is replaced by Stochastic gradient descent algorithm for sample generation on the original set of indicator flags to increase sample data volume. The Faster R-CNN algorithm is used for feature extraction and recognition of pavement signs based on the Faster R-CNN algorithm. The recognition model is fine-tuned based on transfer learning. Actual road environments are used to verify the proposed method. The result shows that the sample images generated by DCGAN effectively amplifies the dataset of the pavement signs,and the recognition accuracy is increased by17.1%. The recognition accuracy in small sample cases increased when sample volume and quality is significantly improved.
作者
程校昭
陈志军
吴超仲
马枫
CHENG Xiaozhao;CHEN Zhijun;WU Chaozhong;MA Feng(Intelligent Transportation Systems Research Center,Wuhan University of Technology,Wuhan 430063)
出处
《交通信息与安全》
CSCD
北大核心
2020年第2期47-54,共8页
Journal of Transport Information and Safety
基金
国家自然科学基金项目(U1764262)资助。