A blind image super-resolution network guided by kernel estimation and structural prior knowledge.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Blind image super-resolution (BISR) aims to restore high-resolution images from degraded low-resolution inputs with unknown kernel parameters.
• Existing methods often focus on kernel estimation but neglect valuable structural prior information within images, especially for textures with high self-similarity.

Purpose of the Study:

• To develop an advanced BISR network that effectively utilizes structural texture prior information.
• To enhance the recovery of high-frequency details and textures in super-resolved images.

Main Methods:

• A two-stage network architecture incorporating a dynamic kernel estimator for degradation embedding.
• A triple path attention mechanism with attention blocks and global feature fusion to extract and leverage structural priors.
• Integration of structural texture as prior knowledge to guide the super-resolution process.

Main Results:

• The proposed method demonstrates superior performance on standard benchmarks (Gaussian8, DIV2KRK) across various degradation types.
• Quantitative and qualitative evaluations confirm improved fidelity and detail recovery compared to state-of-the-art techniques.
• The network successfully addresses limitations in recovering textures with strong self-similarity.

Conclusions:

• The novel BISR network effectively integrates structural prior information, significantly advancing image super-resolution capabilities.
• This approach offers a more robust solution for recovering high-resolution images, particularly those with complex textures.
• Open-source code is available, facilitating further research and application.