基于细节增强和多颜色空间学习的联合监督水下图像增强算法  

Joint supervised underwater image enhancement algorithm based on detail enhancement and multi-color space learning

作　　者：胡锐[1,2] 程家亮 胡伏原 HU Rui;CHENG Jialiang;HU Fuyuan(Information Construction and Management Center,Suzhou University of Science and Technology,Suzhou 215009,China;School of Electronic and Information Engineering,Suzhou University of Science and Technology,Suzhou 215009,China)

机构地区：[1]苏州科技大学信息化建设与管理中心,江苏苏州215009 [2]苏州科技大学电子与信息工程学院,江苏苏州215009

出　　处：《现代电子技术》2025年第1期23-28,共6页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61472267)。

摘　　要：由于水下特殊的成像环境,水下图像往往具有严重的色偏雾化等现象。因此文中根据水下光学成像模型设计了一种新的增强算法,即基于细节增强和多颜色空间学习的无监督水下图像增强算法(UUIE-DEMCSL)。该算法设计了一种基于多颜色空间的增强网络,将输入转换为多个颜色空间(HSV、RGB、LAB)进行特征提取,并将提取到的特征融合,使得网络能学习到更多的图像特征信息,从而对输入图像进行更为精确的增强。最后,UUIE-DEMCSL根据水下光学成像模型和联合监督学习框架进行设计,使其更适合水下图像增强任务的应用场景。在不同数据集上大量的实验结果表明,文中提出的UUIE-DEMCSL算法能生成视觉质量良好的水下增强图像,且各项指标具有显著的优势。Due to the special underwater imaging environment,serious color offset and atomization usually occurs to underwater images.In view of this,on the basis of the underwater optical imaging model,this paper designs a new enhancement algorithm,named unsupervised underwater image enhancement algorithm based on detail enhancement and multi-color space learning(UUIE-DEMCSL).In this algorithm,an enhancement network based on multi-color space is designed.This network converts the input into multiple color spaces(HSV,RGB and LAB)for feature extraction.The extracted features are fused,so that the network can learn more image feature information and enhance the input image more accurately.The UUIE-DEMCSL is designed according to the underwater optical imaging model and the joint supervised learning framework,which is more suitable for the application scenarios of underwater image enhancement tasks.A large number of experimental results on different data sets show that the proposed UUIE-DEMCSL algorithm can generate underwater enhanced images with good visual quality,and each index of the algorithm has significant advantages.

关 键 词：水下图像增强 多颜色空间学习 无监督学习 细节增强 特征提取 特征融合 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391.41[电子电信—信息与通信工程]