基于残差密集块与注意力机制的图像去雾网络
2021-08-02李硕士刘洪瑞甘永东朱新山张军
李硕士 刘洪瑞 甘永东 朱新山 张军



摘 要:基于卷积神经网络的单幅图像去雾算法虽然取得了一定进展,但仍然存在去雾不完全和伪影等问题. 基于这一现状,提出了一种以编码器-解码器结构为基本框架,融合注意力机制与残差密集块的单幅图像去雾网络. 首先,利用网络中的编码器、特征恢复模块和解码器三个部分直接对去雾后的图像进行预测;然后,在网络中引入本文所设计的带有注意力机制的残差密集块,提升网络的特征提取能力;最后,基于注意力機制提出自适应跳跃连接模块,增强网络对去雾图像细节的恢复能力. 实验结果表明,与现有去雾方法相比,提出的去雾网络在合成有雾图像数据集和真实有雾图像上均取得了较为理想的去雾效果.
关键词:图像去雾;深度神经网络;编码器-解码器;注意力机制
中图分类号:TP391.4 文献标志码:A
Image Dehazing Network Based on Residual Dense
Block and Attention Mechanism
LI Shuoshi1,2,LIU Hongrui1,GAN Yongdong1,ZHU Xinshan1,2?,ZHANG Jun1
(1. School of Electrical and Information Engineer,Tianjin University,Tianjin 300072,China;
2. State Key Laboratory of Digital Publishing Technology,Beijing 100871,China)
Abstract:Although the single image dehazing algorithms based on the deep convolutional neural network have made significant progress,there are still some problems, such as poor visibility and artifacts. To overcome these shortcomings,we present a single image dehazing network, taking the encoder-decoder structure as the basic frame and combining the attention mechanism and residual dense block. First,the scheme integrates an encoder, a feature recovery module and the decoder to directly predict the clear images. Then, the residual dense block with attention mechanism is introduced into the dehazing network so as to improve the network's feature extraction ability. Finally, based on the attention mechanism, an adaptive skip connection module is proposed to enhance the network recovering ability for the clear images details. Experimental results show that the proposed dehazing network provides better dehazing results on synthetic datasets and real-world images.
Key words:image dehazing;deep neural networks;encoder-decoder;attention mechanism
在雾、霾等天气下,雾气中悬浮颗粒会对光线的散射造成影响,导致图像传感器所捕捉到的图像出现对比度下降等图像质量退化问题. 这些退化的图像无法真实反应场景中所存在物体的结构、颜色等信息,降低了其在图像分类[1]、目标检测[2]等计算机视觉任务中的应用价值. 因此,图像去雾是目前计算机视觉研究中的一个重要问题.
目前已有的图像去雾方法主要可以分为两类,一类是基于先验知识的去雾方法,另一类是基于深度学习的去雾方法. 基于先验知识的去雾方法通常……
