SCMamba: A Space Correction State Space Model for Image Restoration
Lei Yang, Jingyi Liu, Ze Shi, Caijuan Shi · 2024
Recent advancements in image restoration (IR) have shown that CNNs and Transformers yield impressive results. However, existing IR methods often suffer from limited receptive fields, high computational complexity, leading to insufficient global representation of the restored images and unacceptable computational costs. Inspired by linear complexity and global receptive field of the State Space Model (SSM), we integrate Mamba to IR tasks for learning long-term dependencies with global receptive fields at a linear computational cost. We propose the Space Correction Mamba (SCMamba) framework, which leverages a global receptive field for long-term pixel context modeling to restore more image details. Experiments on the public datasets Set5, Set14 and Urban100 show that SCMamba achieves highly competitive results.