基于全局特征提取和提示学习的水下图像增强  

Underwater image enhancement based on global feature extraction and prompt learning

作　　者：张明华[1] 刘佳艺 石少华 宋巍 ZHANG Minghua;LIU Jiayi;SHI Shaohua;SONG Wei(College of Information,Shanghai Ocean University,Shanghai 201306,China;Ministry of Natural Resources East China Sea Survey Center,Shanghai 200137,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]自然资源部东海调查中心,上海200137

出　　处：《华中科技大学学报(自然科学版)》2025年第3期31-40,共10页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61972240)。

摘　　要：针对水下成像在受到光线、水深和悬浮物干扰等影响时出现的对比度低、信息丢失和色彩失真的退化问题,基于全局特征提取和提示学习提出了一种改进CycleGAN的水下图像增强算法.该方法一方面在网络的编码器中利用全局上下文模块提取图像的全局信息,同时引入注意力机制关注重要信息部分,在下采样的过程中保留更多细节纹理信息;另一方面在解码器中基于提示学习技术设计了一种多尺度退化提示模块,将提取到的多尺度特征中的退化信息编码为易于学习的退化提示,并逐层注入解码器中以引导网络在提高对比度的同时在去噪和去模糊方面也表现得更好.此外,利用全局相似性和感知损失相结合的级联损失函数降低了源域图像与目标域图像的差异.实验采用水下图像质量度量(UIQM)、水下彩色图像质量度量(UCIQE)和自然图像质量度量(NIQE)三种水下图像评估指标,结果表明经本文方法增强后的图像在UIQM和UCIQE上相较于原始的CycleGAN分别提升了13.6%和18.4%,并且综合三个指标值优于其他对比算法,表明本文方法既改善了图像的色彩质量和对比度,同时又保留了原始图像的细节部分,最终经过实验证明了增强后的水下图像的有效性.To address the issues of low contrast,information loss,and color distortion in underwater imaging caused by light interference,water depth,and suspended particles,an improved CycleGAN-based underwater image enhancement algorithm was proposed based on global feature extraction and prompt learning.On the one hand,the global context module was used to extract the global information of the image in the encoder of the network,and the attention mechanism was introduced to pay attention to the important information part,so as to retain more detailed texture information in the process of downsampling.On the other hand,a multi-scale degradation prompt module was designed based on prompt learning technology in the decoder,which was injected layer by layer into the decoder to guide the network to perform better in denoising and deblurring while improving contrast,and the degradation information extracted from multi-scale features was encoded into easily learnable degradation prompts.In addition,the cascading loss function combining global similarity and perceived loss reduces the difference between the images in the low source domain and the target domain.In the experiment,three underwater image evaluation indexes were used:underwater image quality metric(UIQM),underwater color image quality metric(UCIQE)and natural image quality metric(NIQE).The results show that the enhanced images in this paper are 13.6%and 18.4%higher than the original CycleGAN in UIQM and UCIQE,respectively,and the combined values of the three indexes are better than those of other comparison algorithms,indicating that the proposed method improves the color quality and contrast of the images.At the same time,the details of the original image are also retained,and the effectiveness of the enhanced underwater image is finally proved by experiments.

关 键 词：图像增强 生成对抗网络 注意力机制 提示学习 感知损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]