基于预训练扩散模型的两阶段高分辨率图像复原方法

打开文本图片集

关键词：图像复原；扩散模型；预训练模型

中图分类号：TP399 文献标志码：A 文章编号：1001-3695（2025）08-040-2545-07

doi：10.19734/j.issn. 1001-3695.2024. 10.0438

Two-stage high-resolution image restoration method based on pre-trained diffusion model

Xie Yuanyuan，Zhou Fei ，Zhou Zhiyuan，Zhang Yutong （ScholofCommnication&IforationEnginering，CongqingUniersityofPosts&Telecoicatios，Chongqing5，Chna）

Abstract：Pretraineddifusion priors forimagerestorationrelyonpretraineddifusion models to handlevarious clasic mage restoration tasks without fine-tuning.However，currentmethodsare iefficentforhigh-resolutionimages andsuferfromOOD issues.Toadressthese problems，thispaper proposedatwo-stage high-resolution image restoration method basedon pretrained difusionmodels，namedtheC2Fmethod.Inthefirststage，the methodobtainedacoarserestorationresultatthefixed sizeof the pretrained model，ensuring outputconsistency.Inthesecond stage，themethodrestoredtheoriginalresolution usingthecoarseresult asastarting point.Ashorterdifusionprocessignificantlyenhancdrestorationspeedandensuredconsistency.In experiments withfacialrestoration，naturalenvironments，ndtheeclasictasks—repair，colorzation，anddeblurring—the two-stage methodachievestop-levelresultsatanyresolution.For1O24resolution，theproposed methodrequires only 22% of the sampling steps of similar methods，achieving 4.5 times faster speed while avoiding OOD issues.It also reachesthe highestlevelsofPSNRandFIDscores.Thispaperdemonstrates thattheproposed methodrestores high-resolution images much faster than other methods，avoids OOD issues，and produces high-quality results.

Keywords：image restoration；diffusion model;pretrained model

0 引言

在采样、存储、传输和处理过程中，图像常常会受到噪声、模糊、分辨率不足等退化因素的影响，从而降低图像质量，影响后续的图像分析和理解。（剩余16544字）