基于全局元注意力和对比学习的图像去雾  

Image dehazing based on the global meta attentionand contrastive learning

作　　者：许光宇[1] 华健 XU Guangyu;HUA Jian(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《齐鲁工业大学学报》2025年第2期17-25,共9页Journal of Qilu University of Technology

基　　金：国家自然科学基金(61471004);安徽理工大学博士专项基金(ZX942)。

摘　　要：现实中的雾霾复杂多变,数据驱动图像去雾算法无法通过学习所有数据集的数据分布来建立一个泛用的映射模型,且只采用清晰无雾图像作为正样本指导去雾网络的训练,而负样本(有雾图像)的关键信息则被忽略。针对上述问题,提出一种基于全局元注意力和对比学习的图像去雾算法。首先,根据雾在图像中的分布特性设计基于多尺度特征提取和集成的去雾网络。其次,构建全局元注意力模块为多尺度去雾网络提供全局注意力优化,可根据输入的有雾图像自适应地调整网络的映射模型。最后,引入自监督对比学习将去雾结果拉近正样本,而远离负样本。实验结果表明,该算法在合成雾图数据集和真实雾图数据集上都取得了较好地去雾性能,在主观和客观评价方面都优于已有代表性的图像去雾方法。The haze in reality is complex and changeable,and the data-driven image dehazing algorithm cannot establish a universal mapping model by learning the data distribution of all data sets,and only use clear images as positive samples to guide the training of the dehazing network,while the key information of negative samples(hazy images)is ignored.To solve these problems,this paper proposes an image dehazing algorithm based on the global meta attention and contrastive learning.Firstly,according to the distribution characteristics of haze in the image,a dehazing network based on multi-scale feature extraction and integration is designed.Secondly,the global meta attention module is constructed to provide the global attention optimization for the multi-scale dehazing network,and the mapping model of the network can be adaptively adjusted according to the input hazy images.Finally,self-supervised contrastive learning is introduced to draw the dehazing results closer to the positive samples and away from the negative samples.A large number of experimental results show that the algorithm achieves good dehazing performance in both synthetic haze map data sets and real haze map data sets,and is superior to the existing representative image dehazing methods in both subjective and objective evaluation.

关 键 词：图像去雾 深度学习 全局元注意力 对比学习 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]