基于混合注意力的单幅图像去雾算法  

Single Image Dehazing Algorithm Based on Mixed Attention

作　　者：王贺 陈巧莹 WANG He;CHEN Qiaoying(College of Physics and Electronic Engineering,Shanxi University,Taiyuan,030006,China)

机构地区：[1]山西大学物理电子工程学院,太原030006

出　　处：《网络新媒体技术》2024年第6期14-20,共7页Network New Media Technology

基　　金：2022年山西省高等学校科技创新项目(2022L008)。

摘　　要：单幅图像去雾是指将模糊图像经过处理恢复成清晰图像的过程。随着计算机图像处理的快速发展,基于深度学习的去雾算法有了较大进展,但仍存在一些问题,如颜色失真、去雾不完全等。为了解决这些问题,设计了一种混合注意力机制,该机制结合了Transformer注意力、通道注意力和像素注意力,并引入可变形卷积用于特征提取,从而构建一个单幅图像去雾网络。为了得到更好的去雾模型,在RESIDE数据集上对该模型进行训练和调试,取得了较好的实验结果:PSNR指标比FFA-Net提升3.5%,比GCANet提升4.2%;SSIM指标比FFA-Net提升0.21%,比GCANet提升1.1%。Single image dehazing refers to the process of restoring a hazy image to a clear one through image processing.Recently,as the rapid development of computer image processing,deep learning-based fog removal algorithms have made great progress,but there are still some problems,such as color distortion,incomplete fog removal and so on.To solve these problems,a hybrid attention mechanism was designed,combining Transformer attention,channel attention,and pixel attention,and introducing deformable convolution for feature extraction,thereby constructing a single image dehazing network.In order to obtain a better defogging model,the model is trained and debuggable in the data set RESIDE,and good experimental results are obtained:the PSNR metric is improved by 3.5%compared to FFA-Net and by 4.2%compared to GCANet,and SSIM metric is improved by 0.21%compared to FFA-Net and by 1.1%compared to GCANet.

关 键 词：计算机图像处理 深度学习 TRANSFORMER 注意力机制 可变形卷积 RESIDE 数据集 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]