基于生成式对抗网络的道路交通模糊图像增强

PDF下载HTML

Title:

Enhancement of blurred road-traffic images based on generative adversarial network

作者:

曹锦纲, 李金华, 郑顾平

华北电力大学 控制与计算机工程学院，河北 保定 071003

Author(s):

CAO Jin’ gang, LI Jinhua, ZHENG Guping

School of Control and Computer Engineering, North China Electric Power University, Baoding 071003, China

关键词:

图像增强; 道路交通; 运动模糊; 多尺度; 多权重; 残差网络; 神经网络; 生成式对抗网络

Keywords:

enhancement; road traffic; motion blur; multi-scale; multi-weight; residual network; neural network; generated adversarial network

分类号:

TP391

DOI:

10.11992/tis.201903041

摘要:

为了提高道路交通模糊图像增强的质量，进一步促进道路交通管理，针对道路交通场景下的运动模糊图像增强问题，提出了一种基于生成式对抗网络的多尺度多路径学习的模型。首先，选用具有多尺度卷积核的神经网络，对输入的图像进行更细致地特征值提取；其次，将局部残差学习和全局残差学习相结合，采用多路径多权重共享的递归学习，并利用判别网络和生成网络间的对抗训练优化网络参数；最后，实现端到端直接生成图像。实验结果表明：提出的模型可以有效地增强道路交通场景下的运动模糊图像，生成的图像细节更加丰富，具有较好的图像视觉效果。

Abstract:

To improve the quality of blurred road-traffic images and facilitate road traffic management, we propose a multi-scale multi-path learning model based on a generative adversarial network, which solves the problem of enhancing motion-blur images in road traffic scenarios. First, the model selects a neural network with a multi-scale convolution kernel to extract the eigenvalues of the input image in more detail. Then, by combining local and global residual learning techniques and applying recursive learning with multi-path and multi-weight sharing, the model performs adversarial training between discriminant and generating networks to optimize the network parameters. Lastly, an image is generated directly end to end. The experimental results show that the proposed model can effectively enhance motion-blur images in road traffic scenarios, and the details of the generated image are richer and have better visual effects.