南京信息工程大学学报2026，Vol.18Issue(2)：173-182,10.DOI:10.13878/j.cnki.jnuist.20250320004
基于双域特征融合的多尺度运动图像去模糊
Multi-scale motion image deblurring based on dual-domain feature fusion
摘要
Abstract
To address the issue of motion blur in images captured from dynamic scenes,which degrades image quality and causes serious loss of detail information,we propose a multi-scale motion image deblurring method based on dual-domain feature fusion.First,a Dual-Domain Feature Fusion Block(DDFFB)is designed,which employs a two-branch structure to extract spatial-and frequency-domain features from blurred images in parallel,followed by deep fusion of dual-domain features to enhance the model's capability for representing high-frequency details.Next,a Multi-Scale Feature Aggregation Module(MSFAM)is introduced,which utilizes cross-channel self-attention to aggregate the encoded features from different scales of blurred images and dynamically adjusts the weights of feature maps at these scales to enhance model robustness.Furthermore,the training loss function is improved,and the train-ing of the network model is supervised using a joint multi-scale loss function combining content loss,wavelet domain reconstruction loss and edge loss,thereby enhancing the deblurring performance.Comparative experiments conduc-ted on two public datasets,GoPro and HIDE,demonstrate that the proposed method achieves Peak Signal-to-Noise Ratio(PSNR)values of 32.56 dB and 30.76 dB,respectively,surpassing all other compared methods.The results indicate that our approach effectively improves blurring performance and exhibits strong robustness.关键词
运动图像去模糊/双域特征融合/多尺度特征聚合/小波域重构损失Key words
motion image deblurring/dual-domain feature fusion/multi-scale feature aggregation/wavelet domain reconstruction loss分类
信息技术与安全科学引用本文复制引用
吴志强,熊邦书,陈九九,欧巧凤,饶智博,余磊..基于双域特征融合的多尺度运动图像去模糊[J].南京信息工程大学学报,2026,18(2):173-182,10.基金项目
国家自然科学基金(62473187,62365014,62401244) （62473187,62365014,62401244）
江西省职业早期青年科技人才培养项目(20244BCE2091) （20244BCE2091）
