Reproduce “Learning Enriched Features via Selective State Spaces Model for Efficient Image Deblurring” on OpenResearch

About this paper

Hu Gao, Depeng Dang

Researchers from Beijing Normal University introduce ALGNet, an efficient U-shaped network for image deblurring that integrates selective State Space Models (SSMs). The network achieves state-of-the-art deblurring performance on synthetic and real-world datasets, notably outperforming previous SSM-based methods like MambaIR by 0.84 dB PSNR on GoPro while reducing computational costs by 96.1% and enabling 3.1 times faster inference.

Key insights

• A synergistic combination of local (simplified channel attention) and global (selective SSM) feature extraction within the CLGF module effectively captures both fine details and long-range dependencies, overcoming the limitations of each approach when used in isolation.
• Pre-processing and the dedicated local branch successfully mitigate the 'local pixel forgetting' and 'channel redundancy' issues inherent in applying SSMs directly to 2D image data, enabling efficient and accurate pixel-level restoration.
• The lightweight, learnable Feature Aggregation module efficiently fuses local and global information, allowing the network to dynamically adjust the importance of local features for superior detail recovery.

Summary by alphaXiv.