AL-GAN:一种融合注意力机制的轻量级GAN水下图像增强模型
AL-GAN:A lightweight GAN underwater image enhancement incorporating attention mechanisms
冯建新 1韩亚军 1潘成胜 1孙传林 1蔡远航1
作者信息
- 1. 辽宁省通信网络与信息处理重点实验室,辽宁大连 116622;大连大学信息工程学院,辽宁大连 116622
- 折叠
摘要
针对水下图像存在对比度低、颜色失真和现有网络模型推理速度慢等问题,提出一种融合注意力机制的轻量级GAN水下图像增强模型.该模型使用PatchGAN作为判别网络,生成网络在FUnIE-GAN模型基础上,使用轻量化模型MobileNet替换原U-Net特征提取网络中参数量极大的VGG16 模型作为特征提取模块,提取水下退化图像特征,使网络模型参数量减少,加快了网络模型的推理速度.进一步在特征提取模块引入通道和空间注意力机制,增强了网络特征提取能力,达到了增强图像细节的目的.在EUVP数据集上进行实验,结果表明:该方法在处理真实水下图像时有很好的效果.与几种现有方法相比,本文中所提方法能够更好地提升对比度,修正色偏,减少图像细节信息损失,在主客观指标上都优于现有方法.
Abstract
Aiming at the problems of low contrast,color distortion and slow reasoning speed of existing network models in underwater images,this paper proposes a lightweight GAN underwater image enhancement model integrating attention mechanism.This model uses PatchGAN as the discriminant network to generate a network based on the Funi-Gan model.MobileNet is used to replace the VGG16 model with a large number of parameters in the original U-Net feature extraction network as the feature extraction module to extract features of underwater degraded images,thus reducing the number of parameters in the network model.The inference speed of network model is accelerated.Furthermore,channel and spatial attention mechanism are introduced into the feature extraction module to enhance the feature extraction ability of the network and achieve the purpose of enhancing image details.Experiments on EUVP data show that the proposed method is effective in processing real underwater images.Compared with several existing methods,the proposed method can improve contrast better,correct color bias,and reduce the loss of image detail information,and is superior to the existing methods in both subjective and objective indexes.
关键词
水下图像增强/注意力机制/轻量级/生成对抗网络Key words
underwater image enhancement/attention mechanism/lightweight/generation of countermeasure network引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据