摘要
针对失真图像恢复问题,提出一种基于生成对抗网络的端到端图像去雾算法,并重新设计了生成网络和鉴别网络结构。生成器采用密集连接的编码-解码结构,来最大化各层之间的信息流传输。为了保证去雾任务的高分辨率输出,在每个密集块加入平滑模块来实现这一目标。另外加入循环训练约束条件在网络估计过程中,使得去雾效果更加精确。实验结果表明,所提出算法指标表现最好,且具有高鲁棒性、高收敛性和高精确度。
Aiming at the problem of distorted image restoration,an end-to-end image defogging algorithm based on generative adversarial network is proposed,and the structures of generative network and identification network are redesigned. The generator adopts a densely connected encoding decoding structure to maximize the flow of information between layers. In order to ensure the high-resolution output of defogging task,a smoothing module is added to each dense block to achieve this goal. In addition,cyclic training constraints are added to make the defogging effect more accurate in the process of network estimation. The extensive experimental results show that the proposed algorithm has the best performance,and is well in robustness,convergence and accuracy.
作者
孙宏伟
潘峰
罗川
Sun Hongwei;Pan Feng;Luo Chuan(School of Rail Transit,Hope College of Southwest Jiaotong University,Chengdu,Sichuan 610400,China)
出处
《黑龙江工业学院学报（综合版）》
2022年第2期72-77,共6页
Journal of Heilongjiang University of Technology(Comprehensive Edition)
基金
西南交通大学希望学院2021年校级课程思政建设项目“Python程序设计”(项目编号:KCSZ2021054)。
关键词
生成对抗网络
图像去雾
密集连接
平滑扩张模块
循环网络
generative adversarial network
image defogging
dense connection
smooth expansion module
cyclic network
