基于循环生成对抗网络和Transformer的单幅图像去雾算法  

Single Image Dehazing Algorithm Based on Cycle Generative Adversarial Networks and Transformer

作　　者：王博 魏伟波 张为栋 潘振宽 李明[2] 李金函 WANG Bo;WEI Wei-bo;ZHANG Wei-dong;PAN Zhen-kuan;LI Ming;LI Jin-han(College of Computer Science and Technology,Qingdao University,Qingdao 266071,China;College of Computer Science and Technology,Ocean University of China,Qingdao 266100,China)

机构地区：[1]青岛大学计算机科学技术学院,青岛266071 [2]中国海洋大学计算机科学与技术学院,青岛266100

出　　处：《青岛大学学报（自然科学版）》2024年第2期89-97,125,共10页Journal of Qingdao University(Natural Science Edition)

基　　金：山东省自然科学基金(批准号:ZR2020QF033)资助。

摘　　要：针对传统去雾算法在配对数据集上训练时产生过拟合的问题,基于密度和深度分解的非配对图像去雾网络模型,改进了自增强缩放网络。引入Transformer机制,将其与深度卷积神经网络模块深度融合,提出了一种使用未配对数据集训练的基于循环生成对抗网络和Transformer的CT-Nets图像去雾算法;提取输入图像的深度信息和散射系数特征值,利用大气散射模型尽可能恢复不同场景下真实雾的浓度信息,以提高去雾图像主观视觉质量;基于Swin-Transformer设计自增强精化层,以获得精细的细粒度信息,提高模型泛化能力和最终预测图像真实性。实验结果表明,相较于基于密度和深度分解的非配对图像去雾网络模型,CT-Nets图像去雾算法的峰值信噪比和结构相似性分别提升4%和4.1%。Aiming at the problem of overfitting in traditional dehazing algorithms trained on paired datasets,a non-paired image dehazing network model based on density and depth decomposition was improved with a self-enhancing scaling network.Introducing the Transformer mechanism and deeply integrating it with deep convolutional neural networks for network module deep fusion,a CT-Nets image dehazing algorithm based on cycle generative adversarial networks and Transformers trained on unpaired datasets was proposed.The depth information and scattering coefficient eigenvalues of the input image were extracted,and the atmospheric scattering model was used to restore the real fog concentration information in different scenes as much as possible to improve the subjective visual quality of the defogged image.Based on Swin-Transformer,a self-enhancing refinement layer to obtain finer-grained information was designed to improve the generalization ability of the model and the authenticity of the final predicted image.The experimental results show that compared to the dehazing via decomposing transmission map into density and depth network model,the peak signal-to-noise ratio and structural similarity of the CT-Nets image dehazing algorithm are improved by 4%and 4.1%,respectively.

关 键 词：深度学习 单幅图像去雾 自监督网络 循环生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]