摘要
由于扩散模型对数据分布建模的卓越能力,基于扩散模型的图像恢复方法表现出色并获得了广泛关注。针对当前基于扩散模型的图像恢复方法中存在的噪声随机扰动以及数据一致性等问题,提出了一种图像恢复模型rddIRM,在传统的噪声扩散的基础上加入残差扩散,实现双重扩散效果。噪声扩散对待恢复图像的全局分布进行建模和学习,保留图像中的高频信息和细节;残差扩散表示从目标图像退化到输入图像的定向扩散,确定待恢复图像的整体结构和恢复方向。此外,提出将反向过程与数据一致性步骤解耦以及动态自适应调整扩散强度有效地提高了模型的计算效率和图像恢复效果。通过实验证明,rddIRM在各个图像恢复任务中都取得了相对优异的结果。
Due to the ability of diffusion models to effectively model the data distributions,image restoration methods based on diffusion models have demonstrated excellent performance and gain widespread attention.Addressing the issue of random noise perturbation and data consistency,this paper introduced a novel image restoration model,rddIRM,to integrate residual diffusion to achieve a dual diffusion effect.The noise diffusion effectively learned the global distribution of the image,preserving high-frequency details.While the residual diffusion directed the process from the target image to the input,improving structu-ral recovery.Additionally,this paper proposed decouple the reverse process from data consistency steps and dynamically adjus-ted diffusion strength,enhancing computational efficiency and restoration results.Experiments show that rddIRM achieves superior performance across various image restoration tasks.
作者
李鹏
惠健
朱佩佩
Li Peng;Hui Jian;Zhu Peipei(Southwest China Institute of Electronic Technology,Chengdu 610036,China;National Key Laboratory of Complex Aviation System Si-mulaiton,Chengdu 610036,China)
出处
《计算机应用研究》
北大核心
2025年第8期2538-2544,共7页
Application Research of Computers
基金
国家自然科学基金资助项目(62303433)。
关键词
图像恢复
扩散模型
数据一致性
image restoration
diffusion model
data consistency
