Hi-Mamba: Hierarchical Mamba for Efficient Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Transformers excel in low-level vision but face quadratic complexity and limited receptive fields.
• State Space Models (SSMs) offer linear complexity and global receptive fields, yet struggle with long-range dependencies in high-resolution images.
• Existing SSM integration methods for vision tasks suffer from relationship degradation and redundant scanning strategies.

Purpose of the Study:

• To introduce Hi-Mamba, an efficient State Space Model architecture for image super-resolution (SR).
• To address the challenges of long-range dependency learning and redundant scanning in SSM-based vision models.
• To achieve state-of-the-art performance in image super-resolution with a computationally efficient model.

Main Methods:

• Proposed Hi-Mamba architecture featuring a Global Hierarchical Mamba Block (GHMB) for comprehensive token interaction and a global receptive field.
• Incorporated a Direction Alternation Module (DAM) to optimize scanning patterns across different layers, enhancing spatial relationship modeling.
• Developed a single-scan image unfolding strategy to mitigate relationship degradation and long-range forgetting issues.

Main Results:

• Hi-Mamba demonstrated significant performance gains, achieving 0.2-0.27dB PSNR improvements on the Urban100 dataset compared to MambaIRv2 across various scaling factors.
• The lightweight Hi-Mamba model outperformed the SRFormer lightweight model by 0.39dB PSNR for 2x SR.
• The proposed method effectively captures long-range dependencies and improves spatial relationship modeling in image super-resolution tasks.

Conclusions:

• Hi-Mamba presents an effective and efficient solution for image super-resolution by leveraging SSMs.
• The novel architecture successfully addresses key limitations of previous SSM-based approaches in computer vision.
• Hi-Mamba sets a new benchmark for lightweight and high-performance image super-resolution models.