=== SUMMARY ===

Technical Summary: Research on Motion Image Deblurring via Spatial–Frequency Dual-Domain Information Fusion

1. Problem Statement

Motion-induced image blur, caused by relative motion between the camera and the scene during exposure, presents a significant challenge in computer vision. The core difficulty lies in the inherent trade-off between global structure modeling and local detail preservation.

• Spatial-domain methods (e.g., CNNs, Transformers) excel at capturing fine-grained textures and edges but often struggle with limited receptive fields, making it difficult to model long-range dependencies and global blur patterns.
• Frequency-domain methods effectively model global blur patterns and long-range dependencies but frequently fail to maintain spatial consistency in complex motion scenarios, often introducing artifacts such as ringing and color block distortions.

Existing approaches that rely solely on one domain or employ simple fusion strategies fail to simultaneously achieve high-fidelity local detail restoration and accurate global structure reconstruction, particularly in real-world scenarios with complex, non-uniform motion blur.

2. Methodology

The authors propose a Spatial–Frequency Dual-Domain Framework built upon a Generative Adversarial Network (GAN) architecture. The generator follows a U-Net encoder-decoder structure featuring two parallel, balanced branches and a constrained dynamic fusion mechanism.

2.1 Dual-Branch Architecture

The generator embeds two parallel feature processing streams within a U-Net backbone:

1. Spatial Feature Branch (SFB): Utilizes stacked residual blocks with 3×3 convolutions, Batch Normalization, and PReLU activations. This branch focuses on extracting local textures, edges, and structural details directly in the pixel domain.
2. Frequency Feature Branch (FFEB): Transforms input features into the frequency domain via 2D Fast Fourier Transform (FFT). It employs a Learnable Fourier Kernel that adaptively reweights frequency components.
   • The kernel decomposes the complex spectrum into magnitude and phase, modulated by learnable matrices ($W_{amp}$ and $W_{phase}$).
   • To reduce parameters, these matrices are factorized into separable outer products of vectors.
   • The mechanism suppresses low-frequency regions and amplifies high-frequency regions to enhance edges while preserving global structure, followed by an Inverse FFT to return to the spatial domain.

2.2 Dual-Domain Dynamic Fusion Block (DDFB)

A core innovation is the DDFB, which adaptively integrates features from both branches.

• Mechanism: It uses a learnable gating parameter ($\alpha$) to control the fusion ratio between the spatial feature map ($X_1$) and the frequency feature map ($X_2$).
• Constraint: The gating mechanism is explicitly constrained to the range $[0.1, 0.9]$ via a scaling and shifting function ($\alpha = \sigma(gate) \cdot 0.8 + 0.1$). This prevents either branch from dominating the fusion process.
• Training Dynamics: The gate is initialized to 0.5 (equal contribution). During training, if the global mean of the gate deviates significantly from 0.5, a temperature coefficient is adjusted to force the gating curve toward smoothness, ensuring balanced learning without requiring explicit branch-specific labels.

2.3 Additional Modules

• Nested Residual Dense Block (NRDB): Enhances feature reuse and representation capacity through cascaded dense connections without downsampling.
• Bi-Stripe Attention (BSA): Introduced in the upsampling stage to address directionally variant motion blur. It applies multi-head self-attention along horizontal and vertical stripes to recover missing directional structures and model long-range dependencies efficiently.

2.4 Loss Function

The model is trained using a composite loss function:
$$L = \lambda_1 L_{rec} + \lambda_2 L_{prec} + \lambda_3 L_{edge} + \lambda_4 L_{adv}$$
This includes pixel-wise reconstruction loss ($L_{rec}$), perceptual loss ($L_{prec}$) from a pre-trained VGG network, edge consistency loss ($L_{edge}$), and adversarial loss ($L_{adv}$) based on WGAN-GP.

3. Key Contributions

1. Balanced Dual-Branch Analysis: The authors constructed strictly parameter-matched single-branch baselines (Spatial-only and Frequency-only) to systematically analyze representational differences. The study confirms that frequency-domain features are inherently superior for capturing global dependencies but lack spatial consistency, while spatial features preserve local details but struggle with global context.
2. Constrained Dynamic Fusion (DDFB): The proposal of a learnable gating mechanism with explicit constraints that adaptively balances spatial and frequency contributions. This prevents feature dominance and enables pixel-wise adaptive fusion, effectively coordinating global structure reconstruction with local detail preservation.
3. Empirical Validation of Frequency Limitations: Through ablation studies and visual analysis, the paper demonstrates that while frequency branches improve global modeling, they introduce ringing artifacts and color block distortions in complex, heavily blurred regions, justifying the necessity of spatial constraints.

4. Experimental Results

The method was evaluated on synthetic (GoPro) and real-world (RealBlur-J, RealBlur-R, Köhler) datasets.

• Quantitative Performance:

  • GoPro: Achieved 31.08 dB PSNR and 0.915 SSIM.
  • RealBlur-J: Achieved 29.98 dB PSNR and 0.879 SSIM, outperforming several Transformer-based methods (e.g., Restormer, FFTformer) and GAN-based competitors.
  • RealBlur-R: Achieved 35.93 dB PSNR and 0.959 SSIM.
  • Köhler (Spatially Varying Blur): Achieved 22.14 dB PSNR, demonstrating robustness to region-wise varying blur kernels where other methods struggle.

• Qualitative Observations: The method effectively suppresses ringing artifacts and produces visually consistent results with sharper edges in dynamic regions compared to single-domain or simply fused approaches.

• Cross-Task Generalization: Without architectural modifications, the framework was successfully applied to image deraining (Rain1400) and Gaussian denoising (CBSD68), achieving competitive results (28.27 dB PSNR on Rain1400), indicating the generalizability of the spatial-frequency collaboration mechanism.

• Ablation Insights:

  • Simple concatenation of dual branches yielded negligible gains over single branches.
  • The DDFB module provided a superadditive gain of over 1.6 dB PSNR over simple fusion.
  • Information-theoretic analysis showed the DDFB increased the synergistic increment ($\Delta I$) by approximately 23% compared to simple concatenation, confirming effective reduction of redundancy.

5. Significance and Claims

The paper claims that the primary significance of this work lies in providing a principled foundation for cross-domain fusion in image deblurring. By moving beyond architectural design to systematically analyze the mechanisms of frequency-domain modeling, the authors demonstrate that:

1. Frequency-domain features are essential for global dependency modeling but are insufficient alone due to spatial inconsistency.
2. Effective deblurring requires a constrained, dynamic fusion strategy rather than fixed weighting or simple stacking.
3. The proposed Spatial–Frequency Dual-Domain Framework offers a robust solution for complex real-world motion blur, effectively balancing the trade-off between global structure and local detail while mitigating the specific artifacts (ringing, color blocks) associated with frequency-only approaches.

The authors conclude that their approach provides a practical and efficient solution for motion image deblurring, offering a clearer understanding of the strengths and limitations of frequency-domain mechanisms for future multi-domain restoration research. Future work is planned to explore adaptive model simplification to reduce redundancy in dual-branch architectures.