Spatial and Frequency Domain Adaptive Fusion Network for Image Deblurring

TL;DR

The paper tackles image deblurring by proposing SFAFNet, a dual-domain architecture that fuses spatial and frequency information through a gated spatial-frequency feature fusion block (GSFFBlock). The GSFFBlock combines a spatial domain module (NAFBlock-based), a learnable, row-wise FDGM low-pass filter to create adaptive frequency subbands, and a gated fusion module (GFM) with gating (GATE) and cross-attention (CAM) to integrate features. Extensive experiments on GoPro, RealBlur, DPDD, and HIDE show state-of-the-art results in motion and defocus deblurring, with comprehensive ablations confirming the contributions of FDGM, GFM, loss terms, and design choices. The work demonstrates that learning to adaptively separate and fuse spatial and frequency information yields superior restoration performance while offering improved efficiency, highlighting a practical path for robust real-world deblurring.

Abstract

Image deblurring aims to reconstruct a latent sharp image from its corresponding blurred one. Although existing methods have achieved good performance, most of them operate exclusively in either the spatial domain or the frequency domain, rarely exploring solutions that fuse both domains. In this paper, we propose a spatial-frequency domain adaptive fusion network (SFAFNet) to address this limitation. Specifically, we design a gated spatial-frequency domain feature fusion block (GSFFBlock), which consists of three key components: a spatial domain information module, a frequency domain information dynamic generation module (FDGM), and a gated fusion module (GFM). The spatial domain information module employs the NAFBlock to integrate local information. Meanwhile, in the FDGM, we design a learnable low-pass filter that dynamically decomposes features into separate frequency subbands, capturing the image-wide receptive field and enabling the adaptive exploration of global contextual information. Additionally, to facilitate information flow and the learning of complementary representations. In the GFM, we present a gating mechanism (GATE) to re-weight spatial and frequency domain features, which are then fused through the cross-attention mechanism (CAM). Experimental results demonstrate that our SFAFNet performs favorably compared to state-of-the-art approaches on commonly used benchmarks.

Figures (9)

• Figure 1: Visual comparison with MR-VNet MR-VNet and FSNet FSNet. The MR-VNet based on spatial domain often overlooks details. Although the FSNet based on frequency captures details well, it struggles with spatially-variant properties. Our SFAFNet adaptively fuses spatial and frequency features, effectively learning detailed information and spatially-variant structures.
• Figure 2: Computational cost vs. PSNR of models on the GoPro dataset Gopro. Our SFAFNet achieves the SOTA performance.
• Figure 3: (a) Overall architecture of the proposed SFAFNet. (b) Simplified channel attention block (SCABlock) extracts shallow features. (c) Gated spatial-frequency domain feature fusion block (GSFFBlock), which consists of N NAFBlocks chen2022simple, a frequency domain information dynamic generation module (FDGM), and a gated fusion module (GFM). (d) NAFBlock used to extract spatial domain features. (e) FDGM dynamically decompose features into separate frequency subbands.
• Figure 4: Gated fusion module (GFM) contains gating mechanism (GATE) and cross-attention mechanism (CAM).
• Figure 5: Image motion deblurring comparisons on the GoPro dataset Gopro. Our SFAFNet recovers perceptually faithful images.