Core Concepts
State-space model MambaIR improves image restoration with local enhancement and channel attention.
Abstract
The article introduces MambaIR, a state-space model for image restoration that addresses the dilemma between global receptive fields and efficient computation. By enhancing the vanilla Mamba with local pixel similarity and channel attention, MambaIR outperforms SwinIR in image super-resolution tasks. The methodology includes Residual State-Space Blocks (RSSBs) and Vision State-Space Modules to capture long-range dependencies efficiently.
- Introduction:
- Image restoration challenges in computer vision.
- Deep learning models like CNNs and Transformers improve performance.
- Structured State Space Models:
- S4 models balance global receptive field and computational efficiency.
- Mamba adapts S4 for deep networks in various tasks.
- Methodology:
- RSSBs enhance standard Mamba for image restoration.
- VSSMs capture long-range dependencies effectively.
- Loss functions include L1 loss for SR and Charbonnier loss for denoising.
- Experiences:
- Extensive experiments show superior performance of MambaIR over baselines.
- Ablation studies highlight the importance of design choices like RSSB components.
- Comparison on Image Super-resolution:
- Quantitative results show MambaIR outperforming state-of-the-art methods in classic SR tasks.
- Model complexity comparisons demonstrate efficiency similar to SwinIR but with a global receptive field.
- Comparison on Image Denoising:
- Results indicate better performance of MambaIR in Gaussian color image denoising compared to other methods.
- Visual comparisons show clearer edges and natural shapes in denoised images.
- Real-world Image Super-resolution:
- Qualitative comparison demonstrates the robustness of MambaIR in real-world image super-resolution tasks.
- Further questions:
How does the linear complexity of MambaIR impact its scalability?
What are the implications of balancing global receptive fields and computational efficiency?
How can the principles of structured state-space models be applied to other computer vision tasks?
Stats
MambaはSwinIRを上回る0.45dBの性能を示す。
Quotes
"Extensive experiments demonstrate the superiority of our method."
"Mamba-based restoration networks naturally activate more pixels."
"Mamba shares the same recursive form as S4 model."