Learning Enriched Features via Selective State Spaces Model for Efficient Image Deblurring

TL;DR

This work tackles the challenge of efficiently removing blur from high-resolution images by enriching features through a selective state-space–based long-range pathway paired with a local connectivity branch. The authors introduce ALGNet, built from ALGBlocks that combine a global selective state-space module with a local simplified channel-attention branch via CLGF, and a lightweight FA module to optimally fuse local and global cues. Empirical results on GoPro, HIDE, RealBlur, and DPDD demonstrate state-of-the-art PSNR/SSIM and sharper reconstructions, with substantial reductions in computational cost compared to prior methods. The approach offers practical advantages for real-world high-resolution deblurring where both accuracy and efficiency matter.

Abstract

Image deblurring aims to restore a high-quality image from its corresponding blurred. The emergence of CNNs and Transformers has enabled significant progress. However, these methods often face the dilemma between eliminating long-range degradation perturbations and maintaining computational efficiency. While the selective state space model (SSM) shows promise in modeling long-range dependencies with linear complexity, it also encounters challenges such as local pixel forgetting and channel redundancy. To address this issue, we propose an efficient image deblurring network that leverages selective state spaces model to aggregate enriched and accurate features. Specifically, we introduce an aggregate local and global information block (ALGBlock) designed to effectively capture and integrate both local invariant properties and non-local information. The ALGBlock comprises two primary modules: a module for capturing local and global features (CLGF), and a feature aggregation module (FA). The CLGF module is composed of two branches: the global branch captures long-range dependency features via a selective state spaces model, while the local branch employs simplified channel attention to model local connectivity, thereby reducing local pixel forgetting and channel redundancy. In addition, we design a FA module to accentuate the local part by recalibrating the weight during the aggregation of the two branches for restoration. Experimental results demonstrate that the proposed method outperforms state-of-the-art approaches on widely used benchmarks.

Figures (8)

• Figure 1: Computational cost vs. PSNR of models on the GoPro dataset Gopro. Our ALGNet achieve the SOTA performance.
• Figure 2: Overall architecture of ALGNet. (a) ALGNet consists of several ALGBlocks and adopts the multi-input and multi-output strategies for image restoration. (b) ALGBlock comprises two primary modules: a module for capturing local and global features (CLGF), and a feature aggregation module (FA). The CLGF module is composed of two branches: (c) the local branch to model local connectivity, while (d) the global branch captures long-range dependency features.
• Figure 3: (a) Local pixels (highlighted by the red dashed line) are susceptible to being forgotten in the flattened 1D sequence due to the extensive distance. (b) Following guo2024mambair, we apply ReLU and global average pooling to the outputs of the global branch to obtain channel activation values. However, a considerable portion of channels remain inactive, indicating channel redundancy.
• Figure 4: Image motion deblurring comparisons on the GoPro dataset Gopro. Compared to the state-of-the-art methods, our ALGNet excels in restoring sharper and perceptually faithful images.
• Figure 5: Image motion deblurring comparisons on the RealBlur dataset realblurrim_2020_ECCV. Our ALGNet recovers image with clearer details.