面向自动驾驶数据生成的风格迁移网络研究

Research on Style Transfer Network for Autonomous Driving Data Generation

自动驾驶数据集的丰富性是保证基于深度学习的自动驾驶算法鲁棒性和可靠性的关键。当前的自动驾驶数据集在夜晚场景和各类气候、天气条件下的数据量仍十分有限,为满足无人驾驶领域的应用需求,本文中构建了风格迁移网络,可将当前自动驾驶数据集转换为夜晚、雪天等多种形式。该网络采用单编码器-双解码器结构,综合语义分割网络、跳跃连接和多尺度鉴别器等多种手段用于提高图像的生成质量,生成的图像具有良好的视觉效果。用真实数据训练deeplabv3+语义分割网络来评价生成图像的结果表明,本文采用的网络生成图像的平均交并比比基于双编码-双解码结构的两种网络(AugGAN和UNIT)分别提升了2.50%和4.41%。

The data abundance of the autonomous driving dataset is the key to ensuring the robustness and reliability of autonomous driving algorithm based on deep learning,but the amount of data with night scenes and var⁃ious climates and weather conditions in current autonomous driving datasets are still very limited.In order to meet the application needs in the field of unmanned driving,a style transfer network is built,which can convert the cur⁃rent autonomous driving data into various forms such as night and snow,etc.The network adopts a structure of sin⁃gle encoder-dual decoder,combined with various means such as semantic segmentation networks,skip connec⁃tions,and multi-scale discriminators to improve the quality of generated images with good vision effects.Deep⁃labv3+semantic segmentation network trained by real data is used to evaluate the images generated and the results show that the mean intersection over union of the images generated by the network adopted is 2.50 and 4.41 percent⁃age points higher than that generated by AugGAN and UNIT networks with double encoder-double decoder structure respectively.