基于双判别器异构CycleGAN框架下多阶通道注意力校准的室外图像去雾  

Outdoor Image Dehazing Based on Multi-Order Channel Attention Calibration Using a Dual-Discriminator Heterogeneous CycleGAN Framework

作　　者：但志平[1,2] 方帅领 孙航 李晶[3] 万俊 DAN Zhi-ping;FANG Shuai-ling;SUN Hang;LI Jing;WAN Jun(College of Computer and Information Technology,China Three Gorges University,Yichang,Hubei 443000,China;Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang,Hubei 443000,China;School of Computer Science,Wuhan University,Wuhan,Hubei 430072,China;School of Information and Safety Engineering,Zhongnan University of Economics And Law,Wuhan,Hubei 430073,China)

机构地区：[1]三峡大学计算机与信息学院,湖北宜昌443000 [2]三峡大学湖北省水电工程智能视觉监测重点实验室,湖北宜昌443000 [3]武汉大学计算机学院,湖北武汉430072 [4]中南财经政法大学信息与安全工程学院,湖北武汉430073

出　　处：《电子学报》2023年第9期2558-2571,共14页Acta Electronica Sinica

基　　金：国家自然科学基金(No.62002233);湖北省自然科学基金(No.2021CFB004)。

摘　　要：图像去雾是计算机视觉领域中一个经典并具有挑战性的研究方向.近年来,基于深度学习的方法在图像去雾领域取得了显著的成绩.然而,大多数去雾算法依赖于合成配对数据训练网络,由于合成数据与真实有雾数据在分布上存在一定的差距,从而限制了这类去雾方法的实际应用.目前基于CycleGAN网络框架的去雾算法将图像去雾视为一般性图像转换问题,忽视了生成器学习的有效性;此外,在恢复图像时缺乏对于局部区域的探索,构建的网络结构中仅采用一阶通道注意力,忽略了深层次通道相关信息的有效利用.为此,本文提出一种基于双判别器异构CycleGAN框架下多阶通道注意力校准的室外图像去雾算法,该方法主要包含双判别器异构循环框架和多阶通道注意力模块.具体来说,双判别器异构CycleGAN框架通过异构批归一化的生成器和约束生成器局部视野的方式,提升算法的收敛效果和增加局部区域关注.为了进一步挖掘对于图像去雾至关重要的特征通道信息,本文通过引入一阶、二阶特征统计量提出了多阶通道注意力模块,从而提升去雾图像的视觉质量.实验结果表明,在公开合成和真实室外数据集上,本文提出的去雾方法相比现有的8种优秀的去雾算法,取得了最好的客观评价指标和视觉效果.Image dehazing is a classic and challenging research direction in the field of computer vision.In recent years,methods based on deep learning have achieved remarkable achievements in image dehazing.However,most existing dehazing algorithms rely on synthetic paired data training network,which limits the practical application of the dehazing methods,due to the discrepancy in the distribution between the synthetic and real-world foggy images.At present,image dehazing algorithms based on the CycleGAN network framework regard image dehazing as general image transformation,ignoring the effectiveness of generator learning.In addition,these algorithms lacks the exploration of local areas in image restoration,and uses only first-order channel attention in the constructed network,ignoring the effective utilization of deep-level channel-related information.Therefore,this paper proposes a dehazing algorithm for outdoor images based on multi-order channel attention calibration using a dual-discriminator heterogeneous CycleGAN framework,which mainly consists of a dual-discriminator heterogeneous cycle framework and multi-order channel attention module.Specifically,the dual-discriminator heterogeneous CycleGAN framework improves the convergence effect of the algorithm and increases the focus of the local area through the batch normalization generator of the heterogeneous CycleGAN and constraining the generator's local field of view.To further explore the feature channel information that is essential for image dehazing,this study employs a multi-order channel attention module by introducing first-order and second-order feature statistics to improve the visual quality of dehazing images.The results of the experiment show that our proposed method outperforms eight state-of-the-art dehazing algorithms on both synthetic and real-world data sets,regarding the extent of objective evaluation and visual quality.

关 键 词：图像去雾 批归一化 异构CycleGAN 双判别器 多阶通道注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]