一种基于生成对抗网络的图像去雾算法
An Image Dehazing Algorithm Based on Generative Adversarial Networks
李博文 1刘进锋1
作者信息
- 1. 宁夏大学信息工程学院,宁夏 银川 750021
- 折叠
摘要
经现有去雾算法处理的图像存在色彩失真、纹理细节丢失、网格伪影等问题,为此,基于生成对抗网络,提出了一种端到端的去雾算法.该算法将U-Net网络引入生成器模块,采用多尺度卷积和跳跃连接融合不同层级的特征,使用混合空洞卷积模块捕获上下文信息,扩大了感受野并缓解了网格伪影问题;同时通过复合损失函数约束图像边界,提高了去雾图像的精细化质量,有效解决了现有去雾算法存在的问题.所提算法在SOTS数据集上的实验结果表明,去雾图像的客观评价指标和感官效果均优于所有对比算法;在UA-DETRAC数据集上的实验证明,经该算法处理的图像可应用于交通场景的目标检测任务中.
Abstract
In view of the issues such as color distortion,loss of texture details and grid artifacts in the images processed by the existing dehazing algorithms,an end-to-end dehazing algorithm based on generative adversarial networks is proposed.In this algorithm,U-Net is introduced into the generator module,which utilizes multi-scale convolution and skip connections to combine the characteristics of different levels.A hybrid dilated convolution module is employed to capture contextual information,expand the receptive field and alleviate the grid artifacts.Moreover,a composite loss function is utilized to constrain the image boundaries,thereby enhancing the fine quality of the dehazed images and effectively addressing the issues encountered by existing dehazing algorithms.Experimental results of the proposed algorithm on the SOTS data set show that both the objective evaluation metrics and the perceptual quality of the dehazed images outperform all the other compared algorithms.Additionally,experiments on the UA-DETARC data set confirm that the images processed by the proposed algorithm can be applied to the object detection tasks in traffic scenes.
关键词
图像去雾/生成对抗网络/混合空洞卷积/复合损失函数/目标检测Key words
image dehazing/generative adversarial networks/hybrid dilated convolution/composite loss function/ob-ject detection引用本文复制引用
基金项目
宁夏自然科学基金项目(2021AAC03084)
出版年
2024
国家科技期刊平台
NSTL
万方数据