摘要
在图像去模糊领域,细节恢复至关重要且具挑战性。由图像信息在频域中的分布特性可知,低频部分通常承载了图像的结构和大致内容,高频部分则包含了丰富的细节信息。提出一种基于高频恢复和引导的两阶段网络,用于去除模糊并恢复细节。通过八度卷积提取高频特征并送入高频恢复子网络,得到清晰的高频特征。将其与原始模糊图像特征融合,并输入细节恢复子网络,结合局部与全局信息,恢复出清晰图像。实验表明,该方法在多个标准数据集上显著提升了性能,尤其在细节恢复方面表现出色。
In the field of image deblurring,detail recovery is both critical and challenging.Given the distribution characteristics of image information in the frequency domain,the low-frequency components typically carry the structural and general content of the image,whereas the high-frequency components contain rich details.Motivated by this observation,this paper introduces a two-stage network that leverages high-frequency recovery and bootstrapping to mitigate ambiguity and restore details.Initially,octave convolution is employed to extract high-frequency features,which are then fed into the high-frequency recovery subnetwork to obtain clear high-frequency features.Subsequently,these features are fused with the original blurry image features and input into the detail recovery subnetwork,integrating local and global information to reconstruct a clear image.Experimental results demonstrate that the proposed method significantly enhances performance across multiple standard datasets,particularly in terms of detail recovery.
作者
胡波
龚兵兵
李纯熠
胡玲碧
HU Bo;GONG Bingbing;LI Chunyi;HU Lingbi(Chongqing Key Laboratory of Image Cognition,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;School of Software and Artificial Intelligence,Chongqing Institute of Engineering,Chongqing 400056,P.R.China)
出处
《重庆邮电大学学报(自然科学版)》
北大核心
2025年第6期931-939,共9页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
重庆市博士后研究项目专项(2022CQBSHTB2052)
重庆市自然科学基金(CSTB2023NSCQ-BHX0187,CSTB2023N SCQ-LZX0085)
重庆市教育委员会科学技术研究计划(KJQN202401901,KJQN202200638)。
关键词
图像去模糊
两阶段网络
高频恢复
八度卷积
高频引导
image deblurring
two-stage network
high-frequency recovery
octave convolution
high-frequency guidance
