Multi-View Single-Scan Visual State-Space Network for Efficient Image Super-Resolution

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Single Image Super-Resolution (SISR) is crucial for enhancing image quality on resource-constrained edge devices.
• Existing visual state space models offer linear-time modeling but struggle with 2D dependencies, leading to increased computation and memory usage.
• Efficient SISR methods are needed to balance performance with the strict latency and memory budgets of edge computing.

Purpose of the Study:

• To propose an efficient and effective Multi-View Single-Scan Visual State Space Network (MVSSN) for Single Image Super-Resolution (SISR).
• To address the computational and memory inefficiencies of current visual state space models in handling 2D image data.
• To achieve high-quality image reconstruction with minimal parameters and computational cost for edge devices.

Main Methods:

• Developed a Multi-View Single-Scan Visual State Space Network (MVSSN) incorporating a shuffled input stacking module (SISM), a multi-view single-scan block (MSSB), and a multi-scale local feature block (MLFB).
• SISM utilizes shuffled channels and lightweight projection for reduced redundancy.
• MSSB employs alternating scan axes and geometric transforms for efficient cross-layer multi-view context modeling.
• MLFB integrates depthwise convolutions and an MLP for detailed local feature restoration.

Main Results:

• MVSSN demonstrates competitive or superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) on standard SISR benchmarks.
• The proposed network achieves these results with fewer than one million parameters and low Floating Point Operations (FLOPs).
• Evaluations on RealSR datasets and downstream tasks like detection and segmentation confirm MVSSN's efficiency and practical utility.

Conclusions:

• MVSSN offers an efficient solution for Single Image Super-Resolution (SISR) on edge devices, balancing performance with computational constraints.
• The network architecture effectively models 2D dependencies with reduced computational overhead.
• While effective, bicubic-trained models may face domain gaps when applied to unknown real-world image degradations.