基于改进GAN的端到端自动驾驶图像生成方法

基于端到端数据系统的自动驾驶系统对驾驶图像存在巨大需求。为解决一般生成式对抗网络模型在扩充驾驶图像数据集时不稳定及生成图像特征缺乏多样性的问题，研究1种改进网络模型LS-InfoGAN。结合最小二乘对抗损失防止模型梯度消失，并缓解生成器优化矛盾，提升模型训练稳定性。通过最大化生成图像与真实图像间的互信息提升生成器特征学习能力，改善生成图像特征多样性。利用转置卷积层还原图像特征，提升生成图像特征清晰度。以自主构建的模拟驾驶场景中获取的带标签驾驶图像集对模型有效性及其数据集扩充应用效果进行验证。实验分析表明:相比改进前模型，LS-InfoGAN模型的图像生成过程稳定性平均提升35%；使用此模型扩充的数据集进行端到端自动驾驶系统中决策网络的训练能在不采集新图像的情况下将系统决策性能提升1%~2%；建议使用此模型扩充图像数据集时将生成图像数量设置为原始训练集图像数量的1~2倍。 

An Image Generation Method for Automated Driving Based on Improved GAN

There is a huge demand for driving images in the automated driving systems based on end-to-end data system. In order to solve the instability of general generative adversarial network model and the lack of diversity of generated image features when expanding the driving image data set, this work proposed an improved network model, LS-InfoGAN. The least-squares loss is used to prevent the model gradient from disappearing and alleviate the contradiction in the generator during optimization, thereby improving stability of the model. The learning ability of the generator is improved by maximizing mutual information between generated images and actual images, thus improving the diversity of its features. The transposed convolutional layer to restore the image features is used to improve the clarity of the generated image features. The effectiveness and application performance of the model are verified with a labeled image dataset acquired in self-built driving scenes. According to the academic analysis in this study, compared with the model before the improvement, the stability of the image generation process of the LS-InfoGAN model is improved by an average of 35%。Besides, when used for training in the decision network of end-to-end self-driving systems, the augmented dataset can improve the decision performance by 1% to 2% without acquiring new images. The recommended number of generated images is 1 to 2 times the number of original images when the model is used to augment the dataset.