Abstract: Recently, encoder-decoder-based architectures have made significant strides in high-resolution image deblurring tasks. However, for complex dynamic motion blur in high-resolution scenarios, existing methods are mostly limited to single-domain feature extraction or shallow cross-domain concatenation, failing to deeply integrate the structural consistency of the spatial domain with the detail recovery capability of the frequency domain. Moreover, these methods exhibit limitations in receptive field design, which restricts multi-scale feature extraction and leads to a bottleneck in perceiving complex, heterogeneous blur patterns. To address these limitations, a Cross-domain multi-scale Deblurring Network (CDNet) was proposed to jointly model spatial and frequency-domain representations. Firstly, a Cross-Domain Gated Attention (CDGA) module that integrates spatial and frequency domains was designed to extract cross-domain features. Then, a dynamic cross-domain weighting mechanism was equipped for the module to achieve adaptive feature fusion and selectively enhance blur-related cues. Finally, a Multi-Scale Feed-Forward Network (MSFFN) was designed to capture contextual information across different receptive fields, thereby strengthening the integration of local and global blur features. Extensive experimental results on three public datasets, GoPro, HIDE, and RealBlur, demonstrate that compared with the baseline model Restormer (Restoration Transformer), the proposed method improves the Peak Signal-to-Noise Ratio (PSNR) by 0.46 dB, 0.29 dB, and 0.04 dB, and improves the Structural Similarity Index Measure (SSIM) index by 0.007, 0.019, and 0.02, respectively. The experimental results indicate that the proposed method outperforms existing state-of-the-art methods across multiple metrics, exhibiting superior deblurring performance and the ability to restore fine-grained image details more effectively.

Key words: image deblurring, dynamic motion blur, attention mechanism, multi-scale feature extraction, spatial-frequency feature fusion

摘要： 近期，基于编码器-解码器的图像去模糊方法在高分辨率复原任务中取得了显著进展。然而，针对高分辨率场景下复杂的动态运动模糊，现有方法多局限于单一域特征提取或浅层域间拼接，未能深度结合空间域的结构一致性与频域的细节恢复能力；同时，这些方法在感受野设计上存在局限，导致多尺度特征提取受限，依然面临对复杂、异构模糊模式感知不足的瓶颈。为了解决这些局限性，提出一种联合对空间域和频域表示进行建模的跨域多尺度去模糊网络(CDNet)。首先，设计一个融合空间域与频率域的跨域门控注意力（CDGA）模块，以进行跨域特征提取；其次，为该模块配备动态跨域加权机制，从而实现自适应的特征融合以及对模糊相关线索的选择性增强；最后，设计一个多尺度前馈网络（MSFFN），用于捕获跨越不同感受野的上下文信息，进而强化局部与全局模糊特征的整合。在GoPro、HIDE和RealBlur这3个数据集上的大量实验结果表明，与基线模型Restormer(Restoration Transformer)相比，所提方法的峰值信噪比（PSNR）分别提升0.46dB、0.29dB和0.04dB，结构相似度（SSIM）指标分别提升0.007、0.019、0.02。实验结果表明，本文方法在多项指标上均优于现有的先进方法，展现出卓越的去模糊性能，并能更有效地还原细粒度的图像细节。

关键词: 图像去模糊, 动态运动模糊, 注意力机制, 多尺度特征提取, 频率-空间域融合