基于改进CycleGAN的水上图像去雾算法  

Overwater image dehazing algorithm using improved CycleGAN

作　　者：黄超 胡勤友 黄子硕 HUANG Chao;HU Qinyou;HUANG Zishuo(Merchant Marine College,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学商船学院,上海201306

出　　处：《上海海事大学学报》2025年第1期17-22,111,共7页Journal of Shanghai Maritime University

基　　金：国家自然科学基金(52372316)。

摘　　要：雾会使水上拍摄的图像质量下降,导致基于视觉的船舶智能感知系统和水域监控系统受到影响;收集水面上的有雾图像和无雾图像难度较大。针对上述问题,提出一种基于改进循环生成对抗网络(cycle-consistent generative adversarial network,CycleGAN)的水上图像去雾算法。将CycleGAN的生成器模块替换为改进后的门控上下文聚合网络(gated context aggregation network,GCANet),同时使用感知损失从高级语义角度约束图像的生成质量。实验表明:在合成数据集上,所提算法的峰值信噪比和结构相似度分别为25.26和0.9047,相较于对比算法分别提高了13.6%~41.2%和10.9%~17.9%,并在水上真实数据集上展示出了更优的清晰度和色彩真实性。Overwater images often suffer from quality degradation due to fog,affecting vision-based ship intelligent perception systems and water monitoring systems.In addition,it is difficult to collect foggy and fog-free overwater images.To address these problems,an overwater image dehazing algorithm based on the improved cycle-consistent generative adversarial network(CycleGAN)is proposed.The generator module of CycleGAN is replaced with the improved gated context aggregation network(GCANet),and a perceptual loss is introduced to constrain image quality from a high-level semantic perspective.Experimental results on a synthetic dataset show that,the proposed algorithm achieves a peak signal-to-noise ratio of 25.26 and a structural similarity index of 0.9047,which are 13.6%to 41.2%higher and 10.9%to 17.9%higher than those of contrast algorithms,respectively.Furthermore,the algorithm demonstrates superior clarity and color fidelity on a real overwater dataset.

关 键 词：图像去雾 循环生成对抗网络(CycleGAN) 门控上下文聚合网络(GCANet) 感知损失 

分 类 号：U675.79[交通运输工程—船舶及航道工程] TP391.4[交通运输工程—船舶与海洋工程]