摘要
水下图像增强技术能够提升水下图像的质量和可视性,在丰富数字媒体资源、水下探测、水下通信等领域具有重要应用价值。近年来,深度学习方法在水下图像增强方面取得了显著的效果。然而,现有的方法计算复杂度高,限制了它们在计算资源有限的场景中的使用。针对这一问题,提出了一种轻量化的水下图像增强方法,该方法基于跨尺度深度蒸馏特征感知,采用U型网络结构,在保证非线性抽象层级抽取的同时,大幅减少了模型参数量。实验结果表明,所提出方法在视觉效果和客观评价指标上均取得了具有竞争力的结果。
Underwater image enhancement technology has the potential to improve the quality and visibility of underwater images,which has important application value in enriching digital media resources,underwater exploration,underwater communication,and other fields.In recent years,deep learning methods have achieved remarkable results in underwater image enhancement.However,the existing methods suffer from high computational complexity,which limits their application in scenarios with limited computational resources.To address this problem,a lightweight underwater image enhancement method is proposed.The method is based on cross-scale depth distillation feature perception and adopts a U-shaped network structure,which can significantly reduce the model parameter size while ensuring the extraction of nonlinear abstract layers.The experimental results show that the proposed method has achieved competitive results in both visual effects and objective evaluation metrics.
作者
吴晓华
李增禄
许章华
周景春
WU Xiaohua;LI Zenglu;XU Zhanghua;ZHOU Jingchun(School of Art&Design,Sanming University,Sanming 365004,China;Network Technology Center,Sanming University,Sanming 365004,China;Fujian Provincial Key Laboratory of Resources and Environment Monitoring&Sustainable Management and Utilization,Sanming University,Sanming 365004,China;Academy of Geography and Ecological Environment,Fuzhou University,Fuzhou 350108,China;Information Science and Technology College,Dalian Maritime University,Dalian 116026,China)
出处
《大气与环境光学学报》
CAS
CSCD
2024年第3期381-390,共10页
Journal of Atmospheric and Environmental Optics
基金
国家自然科学基金青年基金(62301105)
教育部教师函[2021]13号第二批人工智能助推教师队伍建设试点项目(三明学院)
教育部产学合作协同育人项目(220503924155959)
三明市社科规划项目(22062)。
关键词
水下图像增强
轻量化
跨尺度
蒸馏特征
U型网络
underwater image enhancement
lightweight
cross-scale
distillation feature
U-shaped network
