DCS-RISR: Dynamic channel splitting for efficient real-world image super-resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Real-world image super-resolution (RISR) aims to enhance image quality under complex, unknown degradations.
• Current RISR methods often employ heavy models, limiting deployment on devices with constrained resources.

Purpose of the Study:

• To propose an efficient scheme for Real-world Image Super-Resolution (RISR) suitable for resource-limited devices.
• To develop a method that balances computational cost, parameter count, and image quality metrics.

Main Methods:

• Introduced a Dynamic Channel Splitting (DCS) scheme for efficient RISR, termed DCS-RISR.
• Developed a light degradation prediction network to simulate real-world degradations and generate a channel splitting vector.
• Proposed a learnable octave convolution block to adaptively manage channel splitting scales for different frequency features.
• Incorporated non-local regularization to enhance performance by leveraging patch information from low-resolution (LR) and high-resolution (HR) subspaces.

Main Results:

• DCS-RISR achieves a superior trade-off between computational cost/parameters and performance metrics (PSNR/SSIM).
• The method effectively handles real-world images with varying degradation levels.
• Experiments on benchmark datasets validate the effectiveness and efficiency of the proposed DCS-RISR approach.

Conclusions:

• DCS-RISR offers an efficient solution for real-world image super-resolution.
• The proposed dynamic channel splitting and adaptive convolution significantly reduce computational overhead and memory usage.
• This work enables practical deployment of high-quality image super-resolution on resource-constrained platforms.